Pinned Tweet
🚨 A REMARKABLE STORY ABOUT ELON MUSK’S SECRET GENERALS IN CHINA: THE TWO MEN WHO BUILT THE SHANGHAI MIRACLE 🏆
🌑 In 2018, Tesla had entered its darkest hour. In the United States, severe production crises and low yield rates had put Elon Musk on the hot seat. Wall Street mpany, eyeing it for short-selling, waiting for the inevitable collapse.
🌏 Across the Pacific, the situation was equally dire. In the massive Chinese market—which accounted for nearly half of global new energy vehicle sales—Tesla was struggling to gain traction. Faced with dismal sales of only 120 cars a month, an enraged Musk even considered disbanding the entire Chinese team.
🧱 The market was notoriously difficult to crack. Because all Teslas were imported, the starting price of 499,000 yuan for the Model 3 deterred most consumers. To lower prices, domestic production was essential. However, the premise for foreign companies to produce cars in China was to establish a joint venture—a compromise the maverick Musk was unwilling to make.
🔮 Tesla needed a miracle in China. That miracle would require two distinct phases spearheaded by two very different men: Robin Ren, the diplomat who would unlock the forbidden door, and Tom Zhu, the commander who would build an empire behind it.
PHASE ONE: THE DIPLOMAT AND THE BREAKTHROUGH
🕵️♂️ Secretly, Musk began looking for a "China hand" to navigate the complex political landscape. Robin Ren (Ren Yuxiang), a fellow alumnus of the University of Pennsylvania, had long been on Musk's radar. Since 2012, Musk had repeatedly invited Ren to join Tesla, convinced he was the missing link.
🎓 Initially, Ren was surprised by the olive branch. He admitted he "knew almost nothing about the automotive industry" and found it hard to imagine that, 20 years after graduation, Musk would suddenly ask to have lunch. But Musk was persistent. As the 22nd International Physics Olympiad champion and Musk's former laboratory partner at UPenn, Ren was held in high regard by the CEO, who once noted that Ren was the only classmate whose physics was better than his own.
🤝 Ren finally joined in May 2015 as Vice President of Tesla Asia Pacific. Musk was set on him not just for his intellect, but for his unique leverage: his identity as a native Shanghainese with deep government relations.
🏛️ He became the unsung hero of the Shanghai project. Under Ren's mediation, Musk began frequently meeting with high-ranking Chinese officials. In April 2017, Ren first articulated the crucial argument that a wholly-owned factory "benefits the upgrading of China's automotive industry," persuading officials that Tesla's technology could drive the local supply chain. With theories of technological independence and industrial chain driving effects, Ren slowly loosened the customary domestic joint venture model.
✈️ By February 2018, the plan was ready. Ren flew to the US to report to Musk with a detailed blueprint for the Shanghai factory, including location maps, financing commitments, and transaction terms. Unfortunately, Musk was deep in the "production hell" phase of the Nevada battery factory. When they finally met, Musk didn't even look at Ren’s slides. He just stared at him and asked, "Are we doing this right?"
🚦 Ren was taken aback. He thought the heavy lifting was done, but realized Musk needed reassurance, not data. Giving a firm affirmative answer, Ren secured the green light.
🔓 In April 2018, the breakthrough arrived. The Chinese government lifted foreign ownership restrictions on new energy vehicles, and Ren seized the opportunity. By July, the Shanghai Municipal Government and Tesla signed a memorandum of cooperation. While Shanghai Mayor Ying Yong and Musk unveiled the project publicly, it was Ren who signed the agreement, quietly cementing his pivotal role.
✍️ Ren secured three extremely favorable terms that forcefully broke the established joint venture model. First, he negotiated land concessions, obtaining 860,000 square meters of land in Lingang at a 90% discount from the market price.
⚡ Second, he secured low-interest loans, obtaining credit support totaling over 16 billion yuan with an interest rate of just 3.9%. Third, he ensured rapid approval, taking only half a year from signing the contract to commencing construction. Robin Ren had successfully delivered Tesla's first taste of "China speed."
PHASE TWO: THE COMMANDER AND THE WAR FOR SPEED
🏗️ With the door successfully opened, someone had to walk through it and build. While Ren moved in high-level diplomatic circles, the on-the-ground reality for Tesla China was chaotic.
🔌 In early 2014, the company was struggling with a "charging anxiety" crisis that was killing sales before they could start. Consumers refused to buy electric cars without a reliable network. Into this breach stepped Tom Zhu. Unlike the diplomatic Ren, Zhu was a man of the earth.
🌍 Born in China but educated in New Zealand with an MBA from Duke University, Zhu had cut his teeth managing tough infrastructure projects in Africa. He was used to dust, delays, and difficult environments. He joined Tesla in April 2014 to build the Supercharger network, but his pragmatic, military-style execution caught Musk’s eye immediately. Despite having zero automotive experience, he was put in charge of Tesla’s entire China operation by the end of the year.
⚔️ If securing the land was Robin Ren's victory, building the factory was Tom Zhu's war. The timeline Musk demanded was widely considered impossible: transform a muddy field in Lingang into a world-class vehicle factory in under a year.
🏠 Zhu moved to the front lines. Known for his no-fuss style and often seen wearing a standard-issue Tesla fleece jacket and a buzz cut, Zhu rented a small, government-subsidized apartment just 10 minutes from the construction site. He paid less than 2,000 yuan ($300) a month for rent, purely so he could be the first one in and the last one out.
🚀 Under his watch, "China Speed" became a reality. He orchestrated a 24/7 construction schedule that stunned the industry. In January 2019, the site was dirt. By October 2019—just 10 months later—the factory was complete and starting trial production. It was a miracle of manufacturing engineering that saved Tesla’s cash flow at a critical time.
💰 The results were undeniable. Two years later, the Shanghai factory contributed half of Tesla's global production capacity, and costs were sharply reduced by 65%. Through the Gigafactory, Tesla solved its production and profitability issues in one fell swoop, eventually surpassing a market value of $1 trillion in October 2021.
UNSTOPPABLE: THE MIRACLE OF SHANGHAI
🌟 While Robin Ren left the company in 2020, Zhu’s star continued to rise. His defining moment came in 2022 during the severe Shanghai COVID-19 lockdown. The city was paralyzed, and factories everywhere were shutting down. For Tesla, a halt in Shanghai meant cutting off half its global cash cow.
🛌 Zhu made a decision that mirrored Musk’s own famous "sleeping on the factory floor" days. He implemented a "closed-loop" system, moving into the factory and sleeping on the floor alongside thousands of his workers.
🥣 For over two months, they lived, ate, and worked inside the facility, cut off from the outside world to keep the assembly lines humming. While other automakers flatlined, Zhu’s army kept delivering cars. By 2022, Giga Shanghai was Tesla's primary export hub, producing over 710,000 vehicles that year—more than half of Tesla's global output.
🤠 Musk, who values "hardcore" commitment above all else, saw in Zhu a mirror image of his own relentless drive. In late 2022, when Tesla's Texas and Berlin factories were struggling to ramp up, Musk didn't hire a local expert. He flew Tom Zhu to Austin.
🦺 Zhu arrived with a team of his most loyal lieutenants from Shanghai, famously appearing at the US factories in their signature Tesla visibility vests, ready to instill "China efficiency" into American operations.
🏆 In April 2023, the former project manager who built charging stations was named Senior Vice President of Automotive. Today, Tom Zhu sits at the very pinnacle of Tesla's hierarchy, effectively serving as the global No. 2, overseeing all global production and sales.
☯️ Ultimately, the miracle of Shanghai wasn't just about steel and software; it was about the collision of two distinct forces. Robin Ren was the velvet glove who rewrote the rules of the game, while Tom Zhu was the iron fist who built the arena. One conquered with handshakes, the other with grit. Musk may have provided the vision, but without his Diplomat to open the gate and his Commander to hold the line, the future would have remained just a dream.
🌑 In 2018, Tesla had entered its darkest hour. In the United States, severe production crises and low yield rates had put Elon Musk on the hot seat. Wall Street mpany, eyeing it for short-selling, waiting for the inevitable collapse.
🌏 Across the Pacific, the situation was equally dire. In the massive Chinese market—which accounted for nearly half of global new energy vehicle sales—Tesla was struggling to gain traction. Faced with dismal sales of only 120 cars a month, an enraged Musk even considered disbanding the entire Chinese team.
🧱 The market was notoriously difficult to crack. Because all Teslas were imported, the starting price of 499,000 yuan for the Model 3 deterred most consumers. To lower prices, domestic production was essential. However, the premise for foreign companies to produce cars in China was to establish a joint venture—a compromise the maverick Musk was unwilling to make.
🔮 Tesla needed a miracle in China. That miracle would require two distinct phases spearheaded by two very different men: Robin Ren, the diplomat who would unlock the forbidden door, and Tom Zhu, the commander who would build an empire behind it.
PHASE ONE: THE DIPLOMAT AND THE BREAKTHROUGH
🕵️♂️ Secretly, Musk began looking for a "China hand" to navigate the complex political landscape. Robin Ren (Ren Yuxiang), a fellow alumnus of the University of Pennsylvania, had long been on Musk's radar. Since 2012, Musk had repeatedly invited Ren to join Tesla, convinced he was the missing link.
🎓 Initially, Ren was surprised by the olive branch. He admitted he "knew almost nothing about the automotive industry" and found it hard to imagine that, 20 years after graduation, Musk would suddenly ask to have lunch. But Musk was persistent. As the 22nd International Physics Olympiad champion and Musk's former laboratory partner at UPenn, Ren was held in high regard by the CEO, who once noted that Ren was the only classmate whose physics was better than his own.
🤝 Ren finally joined in May 2015 as Vice President of Tesla Asia Pacific. Musk was set on him not just for his intellect, but for his unique leverage: his identity as a native Shanghainese with deep government relations.
🏛️ He became the unsung hero of the Shanghai project. Under Ren's mediation, Musk began frequently meeting with high-ranking Chinese officials. In April 2017, Ren first articulated the crucial argument that a wholly-owned factory "benefits the upgrading of China's automotive industry," persuading officials that Tesla's technology could drive the local supply chain. With theories of technological independence and industrial chain driving effects, Ren slowly loosened the customary domestic joint venture model.
✈️ By February 2018, the plan was ready. Ren flew to the US to report to Musk with a detailed blueprint for the Shanghai factory, including location maps, financing commitments, and transaction terms. Unfortunately, Musk was deep in the "production hell" phase of the Nevada battery factory. When they finally met, Musk didn't even look at Ren’s slides. He just stared at him and asked, "Are we doing this right?"
🚦 Ren was taken aback. He thought the heavy lifting was done, but realized Musk needed reassurance, not data. Giving a firm affirmative answer, Ren secured the green light.
🔓 In April 2018, the breakthrough arrived. The Chinese government lifted foreign ownership restrictions on new energy vehicles, and Ren seized the opportunity. By July, the Shanghai Municipal Government and Tesla signed a memorandum of cooperation. While Shanghai Mayor Ying Yong and Musk unveiled the project publicly, it was Ren who signed the agreement, quietly cementing his pivotal role.
✍️ Ren secured three extremely favorable terms that forcefully broke the established joint venture model. First, he negotiated land concessions, obtaining 860,000 square meters of land in Lingang at a 90% discount from the market price.
⚡ Second, he secured low-interest loans, obtaining credit support totaling over 16 billion yuan with an interest rate of just 3.9%. Third, he ensured rapid approval, taking only half a year from signing the contract to commencing construction. Robin Ren had successfully delivered Tesla's first taste of "China speed."
PHASE TWO: THE COMMANDER AND THE WAR FOR SPEED
🏗️ With the door successfully opened, someone had to walk through it and build. While Ren moved in high-level diplomatic circles, the on-the-ground reality for Tesla China was chaotic.
🔌 In early 2014, the company was struggling with a "charging anxiety" crisis that was killing sales before they could start. Consumers refused to buy electric cars without a reliable network. Into this breach stepped Tom Zhu. Unlike the diplomatic Ren, Zhu was a man of the earth.
🌍 Born in China but educated in New Zealand with an MBA from Duke University, Zhu had cut his teeth managing tough infrastructure projects in Africa. He was used to dust, delays, and difficult environments. He joined Tesla in April 2014 to build the Supercharger network, but his pragmatic, military-style execution caught Musk’s eye immediately. Despite having zero automotive experience, he was put in charge of Tesla’s entire China operation by the end of the year.
⚔️ If securing the land was Robin Ren's victory, building the factory was Tom Zhu's war. The timeline Musk demanded was widely considered impossible: transform a muddy field in Lingang into a world-class vehicle factory in under a year.
🏠 Zhu moved to the front lines. Known for his no-fuss style and often seen wearing a standard-issue Tesla fleece jacket and a buzz cut, Zhu rented a small, government-subsidized apartment just 10 minutes from the construction site. He paid less than 2,000 yuan ($300) a month for rent, purely so he could be the first one in and the last one out.
🚀 Under his watch, "China Speed" became a reality. He orchestrated a 24/7 construction schedule that stunned the industry. In January 2019, the site was dirt. By October 2019—just 10 months later—the factory was complete and starting trial production. It was a miracle of manufacturing engineering that saved Tesla’s cash flow at a critical time.
💰 The results were undeniable. Two years later, the Shanghai factory contributed half of Tesla's global production capacity, and costs were sharply reduced by 65%. Through the Gigafactory, Tesla solved its production and profitability issues in one fell swoop, eventually surpassing a market value of $1 trillion in October 2021.
UNSTOPPABLE: THE MIRACLE OF SHANGHAI
🌟 While Robin Ren left the company in 2020, Zhu’s star continued to rise. His defining moment came in 2022 during the severe Shanghai COVID-19 lockdown. The city was paralyzed, and factories everywhere were shutting down. For Tesla, a halt in Shanghai meant cutting off half its global cash cow.
🛌 Zhu made a decision that mirrored Musk’s own famous "sleeping on the factory floor" days. He implemented a "closed-loop" system, moving into the factory and sleeping on the floor alongside thousands of his workers.
🥣 For over two months, they lived, ate, and worked inside the facility, cut off from the outside world to keep the assembly lines humming. While other automakers flatlined, Zhu’s army kept delivering cars. By 2022, Giga Shanghai was Tesla's primary export hub, producing over 710,000 vehicles that year—more than half of Tesla's global output.
🤠 Musk, who values "hardcore" commitment above all else, saw in Zhu a mirror image of his own relentless drive. In late 2022, when Tesla's Texas and Berlin factories were struggling to ramp up, Musk didn't hire a local expert. He flew Tom Zhu to Austin.
🦺 Zhu arrived with a team of his most loyal lieutenants from Shanghai, famously appearing at the US factories in their signature Tesla visibility vests, ready to instill "China efficiency" into American operations.
🏆 In April 2023, the former project manager who built charging stations was named Senior Vice President of Automotive. Today, Tom Zhu sits at the very pinnacle of Tesla's hierarchy, effectively serving as the global No. 2, overseeing all global production and sales.
☯️ Ultimately, the miracle of Shanghai wasn't just about steel and software; it was about the collision of two distinct forces. Robin Ren was the velvet glove who rewrote the rules of the game, while Tom Zhu was the iron fist who built the arena. One conquered with handshakes, the other with grit. Musk may have provided the vision, but without his Diplomat to open the gate and his Commander to hold the line, the future would have remained just a dream.
111
404
2,307
278,002
TL;DR ⏳ How can a Tesla drive using nothing but cameras? Four of Tesla's AI inventions answer that, one step at a time.
👀 First, it figures out what's around the car. It turns camera images into a 3D map and marks every spot blocked by something solid, whether it's a car, a person, or a random box in the road.
🛣️ Next, it reads the road itself, including hills, dips, curbs, speed bumps, and painted lines, so it knows where it can actually drive.
🧠 Then it decides what to do. At a busy left turn, it weighs choices like "go now" or "let the pedestrian cross first" and favors the one a good human driver would make.
🅿️ Finally, it parks. It spots empty spaces, avoids no-parking signs and red curbs, and parks itself in the one you tap on the screen.
And none of this is limited to cars. Tesla describes the same tech running a humanoid robot, which is exactly where Optimus comes in.
👀 First, it figures out what's around the car. It turns camera images into a 3D map and marks every spot blocked by something solid, whether it's a car, a person, or a random box in the road.
🛣️ Next, it reads the road itself, including hills, dips, curbs, speed bumps, and painted lines, so it knows where it can actually drive.
🧠 Then it decides what to do. At a busy left turn, it weighs choices like "go now" or "let the pedestrian cross first" and favors the one a good human driver would make.
🅿️ Finally, it parks. It spots empty spaces, avoids no-parking signs and red curbs, and parks itself in the one you tap on the screen.
And none of this is limited to cars. Tesla describes the same tech running a humanoid robot, which is exactly where Optimus comes in.
PURE VISION BREAKTHROUGHS 🚨 Tesla has designed a camera-only brain for the four jobs every self-driving car must master, and cars may be just the beginning 😎
📸 FROM FLAT PIXELS TO A 3D WORLD
Every camera on your Tesla sees the world as a flat picture. Your car, though, has d that is anything but flat, packed with parked cars, sloping roads, painted lines, and pedestrians about to step off the curb. Turning those flat pictures into confident moves, using cameras alone with no LiDAR or radar on the road, is one of the hardest problems in autonomy.
Tesla has been designing its answer piece by piece, and four of its inventions, read together, tell the whole story. They trace the journey from raw camera pixels to a car that understands its surroundings, decides what to do, and parks itself in the exact spot you tapped.
Think about everything you do without even noticing when you drive. You spot the car stopped ahead, feel the road dip into a hill, decide whether to let the pedestrian cross first, and then squeeze into a tight spot without kissing the pillar. Those four instincts are exactly what Tesla is teaching its cameras.
And here's the beautiful part: all four run on the same engine. At the heart of each one sits a single pipeline that turns your car's camera feeds into a living 3D model of the world around it. Each invention then teaches that engine a brand-new skill, so let's break them down one by one.
🧊 JOB ONE: SEEING WHAT'S THERE
The first invention, already officially granted, tackles the most basic question of all. It slices the space around the car into voxels, which are simply 3D pixels, like a giant invisible Minecraft grid wrapped around the vehicle. For every voxel, the AI predicts whether that little cube of space is occupied by something with mass.
The genius here is that the system doesn't need to know what the object is. A car, a pedestrian, or a traffic sign in your path all get the same treatment, because a solid voxel in front of you is something to avoid no matter what it's called. That makes the approach naturally robust against strange, one-off objects that no classifier was ever trained to name.
The pipeline is a work of art. Images from each camera are rectified and compressed into features by image networks (RegNet and BiFPN in Tesla's example). A transformer then lets every 3D location "ask" the camera images which pixels belong to it, fusing all the separate views into one unified 3D space.
Next, the current 3D snapshot is stacked with the previous few, aligned using the car's own trajectory, so the model sees both space and time. Deconvolution layers then expand those fused features back into a full voxel grid.
The outputs go far beyond a simple yes or no. The model can predict occupancy flow, meaning how fast each occupied voxel is moving, plus 3D semantics, such as whether a cluster of voxels is a curb, a building, or a moving vehicle.
The resolution is smart too. Tesla's example uses 33 cm voxels by default but shrinks them to around 10 cm when they're occupied and close to the car or the driving surface. If a curb fills only 16 cm of a 33 cm voxel, a regression step estimates how much of it is actually solid, and interpolation lets software query any point inside a voxel.
So how does a camera-only model learn 3D in the first place? During training, Tesla can use sensor data such as LiDAR to build the answer key, then teach the model to reproduce it from camera footage. Once trained, Tesla is crystal clear on this point: the model predicts 3D occupancy using only the 2D images from the cameras.
The final result is a queryable dataset. Any other piece of software can simply ask "is the point at X, Y, Z occupied?" and get an instant answer.
🛣️ JOB TWO: READING THE ROAD
Knowing where the obstacles are is only half the picture. The second invention turns the AI's attention to the ground the car is actually driving on, which is rarely the perfectly flat plane that simple systems assume.
The model treats the ground as a lattice of points with known X and Y coordinates and predicts the missing piece, the Z value, which is the elevation of each point. Connect those points and you get a 3D mesh of the road itself, revealing hillcrests, downhill grades, banked ramps, speed bumps, and potholes.
Every point carries attributes too. The model can tell whether a surface is drivable, what material it is (grass, dirt, asphalt, or concrete), whether it's road or sidewalk, whether it has paint and what color, and where the curbs are. Picture the car labeling every patch of ground in view with something like "asphalt, one meter above us, drivable, yellow paint."
Training is clever here as well. A 3D point cloud of the surroundings, from LiDAR or triangulated from camera images, gets paired with a segmentation network that labels pixels as road, curb, sidewalk, and so on. Points belonging to the same surface are grouped, a smooth mesh is fitted over them to wipe out noise, and that clean mesh becomes the answer key.
One subtle fix really stands out. When your car rolls over a speed bump, everything measured relative to the car seems to jump, even though the world didn't move at all. So the coordinates are anchored to the road surface instead of the car, which smooths away the jitter caused by the car's own bouncing.
The real magic appears when this invention teams up with the first one. In Tesla's example, occupancy detection covers a vertical window of about 3 meters above and below the car, so a traffic cone sitting on a steep downhill road 4 meters below you would fall outside it.
The surface model solves this by first figuring out that the road there sits 4 meters lower. The cone is then measured against the hill's surface rather than against the car, and the two results are combined. Separating "where is the ground" from "what's on the ground" effectively stretches how far the car can see.
Because all of this is computed from images on the fly, the car can understand a road it has never driven before without first pinpointing itself on a prebuilt map.
🧠 JOB THREE: DECIDING WHO GOES FIRST
A perfect 3D picture means nothing if the car can't decide what to do with it. Tesla calls the third invention joint behavior planning and forecasting, and that name gives the game away.
Tesla's example is a classic: you're turning left into the far lane, a pedestrian is crossing that lane, and another car is turning right onto the same road behind them. It's the same kind of unprotected left that Tesla walked through on stage at AI Day 2022.
The traditional approach would forecast every possible path for the pedestrian and the other car, then grind through an optimization over all of your own possible paths. Tesla points out how much processing power, time, and energy that burns, especially when decisions may be made as often as every 50 milliseconds, or 20 times per second.
Tesla's answer is a hierarchical nodal graph, which works a lot like a choose-your-own-adventure book. At the top sits a goal layer holding what the car wants to achieve, like turning left into the far lane. Below it come layers of interaction nodes, each one a different way the encounter could play out.
The first branch might be "drive in front of the pedestrian" versus "yield and turn behind them." Under the yield branch, the next layer splits again into "slip in ahead of the other car" or "let both pass first." Each node bundles your move and the other road users' moves into one package, so planning and forecasting merge into a single, much simpler question: in what order do we pass each other?
Every node gets a score, which can come from a neural network. The score weighs four things: physics checks such as collisions, ride comfort, how likely a human would be to take over, and a human-like discriminator that judges whether a human driver would make the same choice.
A path's total score combines the scores of its nodes, and the highest-scoring path wins and steers the car.
The efficiency tricks are where this really shines. A node is only expanded when its score clears a threshold, so a branch like "cut in front of the pedestrian" can be dropped on the spot and never explored further. When the graph hits a size cap, the weakest node gets cut, keeping the workload steady instead of spiking when a crowd suddenly shows up.
The fleet feeds this graph too. A server can merge scenarios reported by many cars into one master graph, skip duplicate nodes, ship it to the vehicles, and push updated versions over time. And once the turn is complete, a fresh goal like "drive straight in the new lane" kicks off the next round.
🅿️ JOB FOUR: PARKING WITH INCH-LEVEL PRECISION
The fourth invention takes everything into the tightest spaces of all. A voxel grid that works great on the open road isn't precise enough when you're squeezing into a parking garage, so this one cranks occupancy up to high fidelity.
The key concept is the signed distance field, or SDF. Instead of a simple occupied-or-empty label, each voxel stores a precise number: its distance to the nearest surface. Positive values sit outside an object, negative values sit inside it, and zero lands exactly on the surface, much like contour lines on a hiking map trace the true shape of a mountain instead of just marking where it is.
To turn those numbers into a picture, the rendering works like a 3D printer building layer by layer. It creates a thin horizontal slice, about an inch thick in Tesla's example, starting at road level, coloring the voxels with negative values and leaving the positive ones transparent. Then it builds the next slice up, and the next, and stacks them all into a complete 3D scene.
Tesla shows the difference side by side. A raw point cloud is memory-hungry and noisy, a plain voxel grid looks blocky, and the SDF version captures the fine shapes needed to thread through tight spaces. A pole is drawn from the voxels the model actually sees, rather than by recognizing "pole" and dropping a stock 3D model into the scene.
Since an SDF measures distance to the nearest surface rather than to your car, the system casts virtual rays from camera pixels into the grid to pin down exactly how far away things are. Nearby objects can then be colored red and medium-range ones yellow.
Paint gets the same voxel-level treatment. Instead of hunting for continuous lines, the model checks whether each voxel of ground carries paint, which lets crosswalks and accessible-parking symbols come through along with their color and thickness.
If you've used High Fidelity Park Assist, which Tesla describes as a 3D view showing the shape and proximity of nearby objects, barriers, vehicles, and painted road markings, this will feel very familiar.
Then the parking flow kicks in:
🔍 The car recognizes it has entered a parking area using cues like its slow speed, the map, a sign listing parking prices, or how nearby cars are positioned.
🧱 The high-fidelity model reconstructs the space, and a model trained on labeled parking lots tells real parking spots apart from random empty space.
🚫 Spots get ruled out by "no parking" signs, time-limited signs checked against the current time, red-painted curbs, or an accessible-parking symbol when the driver isn't entitled to use it.
🏆 The survivors get a suitability score that favors spots closer to you and requiring fewer turns and fewer switches between drive and reverse, and you can even ask to park away from other cars.
👆 The top spot gets a "P" marker, you tap the one you want, confirm, and the car parks itself while the screen tracks its progress.
Scores keep updating as you drive, so a spot that looked awkward can jump to the top after you make a turn. Because paint alone can define a spot, it works in a wide-open stadium lot with no curbs in sight, where Tesla notes that ultrasonic-only systems fall short because they can't see paint at all.
It handles both perpendicular and parallel spots, and anyone who has used tap-to-park Autopark will recognize that "P" icon instantly.
🧩 HOW THE PIECES LOCK TOGETHER
Lay all four side by side and the architecture is gorgeous. Occupancy says what's solid and what's moving, the surface model describes the ground under everything, the planner turns it all into a decision, and the high-fidelity layer adds inch-level precision for tight spaces. Tesla even names occupancy as a way for the planner to detect the road users it reasons about, so the links are built right in.
They share the same DNA as well. Cameras go in, 3D understanding comes out, and data from the fleet keeps flowing back to improve the models.
The sharpest owners in the replies will raise a fair point here: FSD went end-to-end with v12, replacing much of the hand-written planning code with a single neural network. That's true, and three of these four inventions date back to 2022, well before that switch.
My take is that the ideas didn't disappear; they got absorbed. Ashok Elluswamy, Tesla's AI software chief and a named inventor on three of the four, has described today's FSD as one network running from camera video straight to control, while noting it can still be probed for its 3D understanding. And the camera-based parking visuals owners use today line up closely with what the fourth invention describes.
🧠 ONE BRAIN, MANY BODIES
Here's where it gets truly exciting. Look closely, and none of these four jobs is really about cars. Seeing what's there, reading the surface underfoot, deciding who goes first, and settling precisely into a spot are the same four jobs facing any machine that moves through our world.
On the road, that's exactly the skill set a Robotaxi needs when nobody is sitting in the driver's seat. But Tesla is clearly thinking bigger. Right alongside cars and trucks, it describes a general-purpose bipedal humanoid robot as one of the machines that can run this technology. Tesla's AI software chief has said the same foundational model behind FSD is being adapted for Optimus.
Picture what that means. For a car, reading the ground means spotting a curb or a hill. For a humanoid robot, the same skill could mean handling a staircase or a ramp. Deciding who goes first at an intersection becomes deciding who goes first in a narrow hallway. Parking becomes walking back to a charging station and docking, which Tesla already sketches out for a robot moving through an office full of desks, chairs, and people.
Your Tesla may be the first body this camera-only brain learned to control. It's very unlikely to be the last.
2
2
15
1,836
HIDDEN GEM 💎 @Tesla has patented a wafer-thin GPS ear that hears satellites in stereo 🦻
Before your Tesla can roll across a parking lot to meet you using Summon, it has to answer a deceptively simple question: Where exactly am I?
Cameras handle the nearby surroundings, but s from satellites thousands of miles overhead. Those signals don't always arrive clean because they can bounce off nearby buildings, reaching the car twice: once directly from the sky and once as an echo. An antenna tuned to just a single frequency can get fooled by the duplicate.
The sky already provides a fix. Navigation satellites broadcast across several frequency bands simultaneously, including L1, L2, and L5, and an antenna capable of catching two of them can cross-check one signal against the other.
That second band unlocks major upgrades. It allows the receiver to reject echoes and correct errors introduced by the atmosphere. It also powers high-precision carrier-phase calculations, pinning down the vehicle's position more accurately than any single band could.
The problem is the hardware. Multi-band antennas have often relied on stacked patches or helix coils with complex feed networks, and they tend to come out thick or limited to a narrow slice of spectrum. A car designed with flush antennas hidden inside the smooth envelope of its body leaves almost no room for hardware like that.
Then there is the reflector. Flat antennas are typically paired with a metal reflector to bounce signals back toward the sky. However, standard metal flips the phase of each wave upside down upon reflection. To ensure the reflected wave returns in phase, it requires a quarter-wavelength gap, which translates to roughly 5 to 6.5 centimeters (about 2 to 2.5 inches) of physical depth at these frequencies.
Tesla flattens the whole problem. It etches a pinwheel pattern of slots into the top copper layer of a printed circuit board, and those cutouts become the antenna itself. A slot in copper captures radio waves much like a standard wire, but it sits completely flush and adds virtually no height.
Each slot originates as a short spoke near the center, spaced evenly like numbers on a clock face. From there, each spoke sweeps outward into a long arm curling around the middle, resembling a spinning galaxy. Together, the slots, circuit board, and underlying feeds can measure anywhere from about 0.5 to 3 millimeters thick.
Next, the pinwheel divides the workload. One spiral feed on the underside connects to the inner spokes, while a second connects to the outer arms. Each feed can be tuned to a distinct band, such as L1 on one feed and L2 or L5 on the other, allowing the antenna to monitor two channels at once.
Both channels route into a diplexer, which acts like a frequency mail sorter. It isolates each band separately before combining them into a single stream for a low-noise amplifier, which boosts the faint signal before it reaches the receiver. Thanks to the symmetrical layout, each feed operates independently, letting Tesla use fewer amplifier parts and reduce costs.
The real trick hides behind the board. Instead of using plain metal, Tesla backs the antenna with an artificial magnetic conductor, a surface tiled with metal patches connected to a lower copper layer through conductive posts called vias. At target frequencies, this surface acts like a specialized mirror, reflecting radio waves without reversing their phase.
This changes the physical spacing requirements completely. When a wave returns in phase, it behaves like a perfectly timed push on a swing, adding energy rather than counteracting it. Because the magnetic conductor provides that timing right at the surface, there is no need for a quarter-wavelength gap, allowing the entire antenna assembly to come in at 15 millimeters (about 0.6 inches) or less in some versions.
The magnetic reflector offers an additional benefit. Its patch dimensions are tailored to create a band gap near target frequencies such as L2 and L5. Gain plots reveal that an antenna backed by this surface picks up a secondary operating band around 1.15 to 1.25 GHz.
An identical antenna mounted over a standard metal disk cannot reach that secondary band. The magnetic mirror does far more than conserve space; it actively helps capture that second signal from above.
Tesla can extend this design even further. In one configuration, the artificial magnetic conductor is integrated onto the same board surrounding the pinwheel, placing the antenna and reflector on a single layer. In another, an existing surface inside the vehicle serves as the reflector, such as a metal-coated housing inside the rearview mirror assembly.
The windshield glass can also play a role. Since the antenna can be positioned directly beneath glass, Tesla can integrate a heating element into the glass above it to clear ice and snow. That heating grid can couple capacitively with the antenna, acting as a parasitic element to boost gain so that even the de-icer contributes to radio performance.
Back in that parking lot, your Tesla still needs to determine its exact coordinates. With this compact design, it can answer the question by listening to dual bands simultaneously and filtering real signals from reflections, all using a pinwheel pattern just a few millimeters thick.
Stereo satellite hearing often meant bulky hardware. Tesla's version can fold it into a sliver of circuit board and a mirror that doesn't flip. The sky speaks in two voices, and this wafer-thin ear can catch both.
Before your Tesla can roll across a parking lot to meet you using Summon, it has to answer a deceptively simple question: Where exactly am I?
Cameras handle the nearby surroundings, but s from satellites thousands of miles overhead. Those signals don't always arrive clean because they can bounce off nearby buildings, reaching the car twice: once directly from the sky and once as an echo. An antenna tuned to just a single frequency can get fooled by the duplicate.
The sky already provides a fix. Navigation satellites broadcast across several frequency bands simultaneously, including L1, L2, and L5, and an antenna capable of catching two of them can cross-check one signal against the other.
That second band unlocks major upgrades. It allows the receiver to reject echoes and correct errors introduced by the atmosphere. It also powers high-precision carrier-phase calculations, pinning down the vehicle's position more accurately than any single band could.
The problem is the hardware. Multi-band antennas have often relied on stacked patches or helix coils with complex feed networks, and they tend to come out thick or limited to a narrow slice of spectrum. A car designed with flush antennas hidden inside the smooth envelope of its body leaves almost no room for hardware like that.
Then there is the reflector. Flat antennas are typically paired with a metal reflector to bounce signals back toward the sky. However, standard metal flips the phase of each wave upside down upon reflection. To ensure the reflected wave returns in phase, it requires a quarter-wavelength gap, which translates to roughly 5 to 6.5 centimeters (about 2 to 2.5 inches) of physical depth at these frequencies.
Tesla flattens the whole problem. It etches a pinwheel pattern of slots into the top copper layer of a printed circuit board, and those cutouts become the antenna itself. A slot in copper captures radio waves much like a standard wire, but it sits completely flush and adds virtually no height.
Each slot originates as a short spoke near the center, spaced evenly like numbers on a clock face. From there, each spoke sweeps outward into a long arm curling around the middle, resembling a spinning galaxy. Together, the slots, circuit board, and underlying feeds can measure anywhere from about 0.5 to 3 millimeters thick.
Next, the pinwheel divides the workload. One spiral feed on the underside connects to the inner spokes, while a second connects to the outer arms. Each feed can be tuned to a distinct band, such as L1 on one feed and L2 or L5 on the other, allowing the antenna to monitor two channels at once.
Both channels route into a diplexer, which acts like a frequency mail sorter. It isolates each band separately before combining them into a single stream for a low-noise amplifier, which boosts the faint signal before it reaches the receiver. Thanks to the symmetrical layout, each feed operates independently, letting Tesla use fewer amplifier parts and reduce costs.
The real trick hides behind the board. Instead of using plain metal, Tesla backs the antenna with an artificial magnetic conductor, a surface tiled with metal patches connected to a lower copper layer through conductive posts called vias. At target frequencies, this surface acts like a specialized mirror, reflecting radio waves without reversing their phase.
This changes the physical spacing requirements completely. When a wave returns in phase, it behaves like a perfectly timed push on a swing, adding energy rather than counteracting it. Because the magnetic conductor provides that timing right at the surface, there is no need for a quarter-wavelength gap, allowing the entire antenna assembly to come in at 15 millimeters (about 0.6 inches) or less in some versions.
The magnetic reflector offers an additional benefit. Its patch dimensions are tailored to create a band gap near target frequencies such as L2 and L5. Gain plots reveal that an antenna backed by this surface picks up a secondary operating band around 1.15 to 1.25 GHz.
An identical antenna mounted over a standard metal disk cannot reach that secondary band. The magnetic mirror does far more than conserve space; it actively helps capture that second signal from above.
Tesla can extend this design even further. In one configuration, the artificial magnetic conductor is integrated onto the same board surrounding the pinwheel, placing the antenna and reflector on a single layer. In another, an existing surface inside the vehicle serves as the reflector, such as a metal-coated housing inside the rearview mirror assembly.
The windshield glass can also play a role. Since the antenna can be positioned directly beneath glass, Tesla can integrate a heating element into the glass above it to clear ice and snow. That heating grid can couple capacitively with the antenna, acting as a parasitic element to boost gain so that even the de-icer contributes to radio performance.
Back in that parking lot, your Tesla still needs to determine its exact coordinates. With this compact design, it can answer the question by listening to dual bands simultaneously and filtering real signals from reflections, all using a pinwheel pattern just a few millimeters thick.
Stereo satellite hearing often meant bulky hardware. Tesla's version can fold it into a sliver of circuit board and a mirror that doesn't flip. The sky speaks in two voices, and this wafer-thin ear can catch both.
7
7
59
3,835
HIDDEN GEM 💎 Tesla has patented a lightning-fast assembly line that helps self-driving AI make sense of what the cameras see 👁️
Every camera frame a Tesla sees becomes a mountain of arithmetic before the car recognizes a single stop sign. Most of the spotlight goes to the r question is what happens to every answer the instant that multiplying ends.
Each layer of a neural network moves in two beats. First comes a huge batch of multiply-and-add work, the dot products that measure how strongly each patch of the image matches a learned pattern. Then every single result needs follow-up treatment, such as an activation function like ReLU that zeroes out negative values, before the next layer can use it.
That second beat is where general-purpose hardware struggles. A traditional processor runs each operation very quickly but only touches a small amount of data at a time. A GPU spreads the work across many cores, yet each core runs on its own instructions, and neither design was originally built for machine learning.
The answers also come out bloated. In one example, the multiplier grid keeps each result in four bytes so it can add up many products without overflowing, even though every number going in was a single byte. Those fat results have to slim back down before the next layer can take them.
Tesla tackles this with a dedicated assembly line bolted right onto the exit of the multiplier grid. Picture that grid as a field of tiny calculators, 96 wide by 96 deep in one example, each one multiplying and accumulating in lockstep. Tesla groups those calculators into lanes, and each lane feeds its own conveyor belt.
Every lane ends at its own processing element inside a block called the vector engine. When the grid finishes, each belt carries its result straight to the workstation waiting at the end of it. In one version, all 96 handoffs travel together over a channel 96 elements wide, with 32 bits per element.
The vector engine doesn't give each station separate orders. One instruction goes out, and all 96 processing elements carry it out at the same moment on their own numbers. It works like a single call from the foreman, with every station on the floor moving as one.
That one instruction carries a lot. Tesla bundles a load, a math operation, and a store into a single 128-bit instruction triad, so one call can tell every lane to grab a value, crunch it, and write a result out. When a step isn't needed, a no-op simply holds its seat.
The instructions also overlap, the way an assembly line keeps several products moving at once. While one instruction stores, the next can do its math and a third can load fresh data. In one version, the engine finishes a full instruction every clock cycle.
Now those bloated answers get their diet. The engine can apply a non-linear function like ReLU and quantize the result in the very same instruction, turning 96 four-byte answers into 96 one-byte values. That output matches the width the multiplier grid takes in, one byte per lane across 96 lanes.
The slimmed results then flow downstream. A post-processing unit can pool them, keeping the strongest signal in each neighborhood, and format them for the trip back to memory.
Underneath all of this, the sharpest trick lives in the engine's storage. Picture each lane's four-byte result as a carton of four eggs.
A byte register such as RB0 holds the first egg from every lane's carton. A half-word register such as RW1 holds the third and fourth eggs from every lane. A full-word register such as RD0 holds the whole carton.
Every register always takes the same slice from each lane, so the slices line up perfectly across all 96. In one example, eight registers that are 32 bits wide in every lane double as sixteen 16-bit registers and thirty-two 8-bit ones. The same physical storage wears three sets of labels, and the instruction decides which label to read.
That lets one bank of storage serve 8-bit, 16-bit, and 32-bit math, with the register size matched to the operation, the data type, or the load instruction. A 96-bit mask works like a row of 96 light switches, deciding exactly which lanes receive new data. Whatever size the next step wants, the right slice is already lined up and waiting.
Now picture that camera frame again. The grid finishes its dot products, and 96 belts deliver their results to 96 waiting stations at the same moment. One instruction applies the activation, shrinks every answer, and sends it on its way, while the next instruction is already loading.
It's tempting to think the multiplying is the whole job. Tesla's design turns everything that comes after it into an assembly line of its own. On this line, a single instruction can finish 96 answers at once.
Every camera frame a Tesla sees becomes a mountain of arithmetic before the car recognizes a single stop sign. Most of the spotlight goes to the r question is what happens to every answer the instant that multiplying ends.
Each layer of a neural network moves in two beats. First comes a huge batch of multiply-and-add work, the dot products that measure how strongly each patch of the image matches a learned pattern. Then every single result needs follow-up treatment, such as an activation function like ReLU that zeroes out negative values, before the next layer can use it.
That second beat is where general-purpose hardware struggles. A traditional processor runs each operation very quickly but only touches a small amount of data at a time. A GPU spreads the work across many cores, yet each core runs on its own instructions, and neither design was originally built for machine learning.
The answers also come out bloated. In one example, the multiplier grid keeps each result in four bytes so it can add up many products without overflowing, even though every number going in was a single byte. Those fat results have to slim back down before the next layer can take them.
Tesla tackles this with a dedicated assembly line bolted right onto the exit of the multiplier grid. Picture that grid as a field of tiny calculators, 96 wide by 96 deep in one example, each one multiplying and accumulating in lockstep. Tesla groups those calculators into lanes, and each lane feeds its own conveyor belt.
Every lane ends at its own processing element inside a block called the vector engine. When the grid finishes, each belt carries its result straight to the workstation waiting at the end of it. In one version, all 96 handoffs travel together over a channel 96 elements wide, with 32 bits per element.
The vector engine doesn't give each station separate orders. One instruction goes out, and all 96 processing elements carry it out at the same moment on their own numbers. It works like a single call from the foreman, with every station on the floor moving as one.
That one instruction carries a lot. Tesla bundles a load, a math operation, and a store into a single 128-bit instruction triad, so one call can tell every lane to grab a value, crunch it, and write a result out. When a step isn't needed, a no-op simply holds its seat.
The instructions also overlap, the way an assembly line keeps several products moving at once. While one instruction stores, the next can do its math and a third can load fresh data. In one version, the engine finishes a full instruction every clock cycle.
Now those bloated answers get their diet. The engine can apply a non-linear function like ReLU and quantize the result in the very same instruction, turning 96 four-byte answers into 96 one-byte values. That output matches the width the multiplier grid takes in, one byte per lane across 96 lanes.
The slimmed results then flow downstream. A post-processing unit can pool them, keeping the strongest signal in each neighborhood, and format them for the trip back to memory.
Underneath all of this, the sharpest trick lives in the engine's storage. Picture each lane's four-byte result as a carton of four eggs.
A byte register such as RB0 holds the first egg from every lane's carton. A half-word register such as RW1 holds the third and fourth eggs from every lane. A full-word register such as RD0 holds the whole carton.
Every register always takes the same slice from each lane, so the slices line up perfectly across all 96. In one example, eight registers that are 32 bits wide in every lane double as sixteen 16-bit registers and thirty-two 8-bit ones. The same physical storage wears three sets of labels, and the instruction decides which label to read.
That lets one bank of storage serve 8-bit, 16-bit, and 32-bit math, with the register size matched to the operation, the data type, or the load instruction. A 96-bit mask works like a row of 96 light switches, deciding exactly which lanes receive new data. Whatever size the next step wants, the right slice is already lined up and waiting.
Now picture that camera frame again. The grid finishes its dot products, and 96 belts deliver their results to 96 waiting stations at the same moment. One instruction applies the activation, shrinks every answer, and sends it on its way, while the next instruction is already loading.
It's tempting to think the multiplying is the whole job. Tesla's design turns everything that comes after it into an assembly line of its own. On this line, a single instruction can finish 96 answers at once.
8
5
35
3,437
BOOM 🚨 Tesla is on the direct verge of crossing 1,000 registered autonomous vehicles in Texas alone 🔥
The latest DMV registrations show another 32 Cybercabs added to the state registry, officially pushing the total dedicated Cybercab count to 463 vehicles 🆒
Combined with 507 sla's active Robotaxi fleet in Texas now stands at 970 vehicles 😎
And steering wheel-less Cybercabs are making up nearly 48% of the commercial fleet at the moment 🔥
The latest DMV registrations show another 32 Cybercabs added to the state registry, officially pushing the total dedicated Cybercab count to 463 vehicles 🆒
Combined with 507 sla's active Robotaxi fleet in Texas now stands at 970 vehicles 😎
And steering wheel-less Cybercabs are making up nearly 48% of the commercial fleet at the moment 🔥
GOOD NEWS 🚖 According to the latest Texas DMV filings, Tesla has registered an additional 11 Cybercabs, pushing the state's total registered Cybercab count to 431 vehicles, accounting for nearly 46% of the state's entire autonomous fleet 🔥
With 507 commercial Model Ys ide them, Tesla's total commercial Robotaxi fleet in Texas has expanded to 938 vehicles 🔥
The total fleet rapidly closes in on the 1,000-vehicle mark 🔥
8
27
249
7,213
If you are seeing third-party trackers report 465 Cybercabs instead of 463, here is why 👇
Back when Tesla first began road-testing Cybercabs in Texas, two early pre-production prototype units were registered under a secondary test permit account before the main Tesla Robotaxi LLC entity page went live.
Our data tracks the primary commercial fleet listed directly under Tesla Robotaxi LLC.
Today's 43 new registrations bring that official commercial total to 463 Cybercabs!
Back when Tesla first began road-testing Cybercabs in Texas, two early pre-production prototype units were registered under a secondary test permit account before the main Tesla Robotaxi LLC entity page went live.
Our data tracks the primary commercial fleet listed directly under Tesla Robotaxi LLC.
Today's 43 new registrations bring that official commercial total to 463 Cybercabs!
1
2
2
1,110
PURE VISION BREAKTHROUGHS 🚨 Tesla has designed a camera-only brain for the four jobs every self-driving car must master, and cars may be just the beginning 😎
📸 FROM FLAT PIXELS TO A 3D WORLD
Every camera on your Tesla sees the world as a flat picture. Your car, though, has d that is anything but flat, packed with parked cars, sloping roads, painted lines, and pedestrians about to step off the curb. Turning those flat pictures into confident moves, using cameras alone with no LiDAR or radar on the road, is one of the hardest problems in autonomy.
Tesla has been designing its answer piece by piece, and four of its inventions, read together, tell the whole story. They trace the journey from raw camera pixels to a car that understands its surroundings, decides what to do, and parks itself in the exact spot you tapped.
Think about everything you do without even noticing when you drive. You spot the car stopped ahead, feel the road dip into a hill, decide whether to let the pedestrian cross first, and then squeeze into a tight spot without kissing the pillar. Those four instincts are exactly what Tesla is teaching its cameras.
And here's the beautiful part: all four run on the same engine. At the heart of each one sits a single pipeline that turns your car's camera feeds into a living 3D model of the world around it. Each invention then teaches that engine a brand-new skill, so let's break them down one by one.
🧊 JOB ONE: SEEING WHAT'S THERE
The first invention, already officially granted, tackles the most basic question of all. It slices the space around the car into voxels, which are simply 3D pixels, like a giant invisible Minecraft grid wrapped around the vehicle. For every voxel, the AI predicts whether that little cube of space is occupied by something with mass.
The genius here is that the system doesn't need to know what the object is. A car, a pedestrian, or a traffic sign in your path all get the same treatment, because a solid voxel in front of you is something to avoid no matter what it's called. That makes the approach naturally robust against strange, one-off objects that no classifier was ever trained to name.
The pipeline is a work of art. Images from each camera are rectified and compressed into features by image networks (RegNet and BiFPN in Tesla's example). A transformer then lets every 3D location "ask" the camera images which pixels belong to it, fusing all the separate views into one unified 3D space.
Next, the current 3D snapshot is stacked with the previous few, aligned using the car's own trajectory, so the model sees both space and time. Deconvolution layers then expand those fused features back into a full voxel grid.
The outputs go far beyond a simple yes or no. The model can predict occupancy flow, meaning how fast each occupied voxel is moving, plus 3D semantics, such as whether a cluster of voxels is a curb, a building, or a moving vehicle.
The resolution is smart too. Tesla's example uses 33 cm voxels by default but shrinks them to around 10 cm when they're occupied and close to the car or the driving surface. If a curb fills only 16 cm of a 33 cm voxel, a regression step estimates how much of it is actually solid, and interpolation lets software query any point inside a voxel.
So how does a camera-only model learn 3D in the first place? During training, Tesla can use sensor data such as LiDAR to build the answer key, then teach the model to reproduce it from camera footage. Once trained, Tesla is crystal clear on this point: the model predicts 3D occupancy using only the 2D images from the cameras.
The final result is a queryable dataset. Any other piece of software can simply ask "is the point at X, Y, Z occupied?" and get an instant answer.
🛣️ JOB TWO: READING THE ROAD
Knowing where the obstacles are is only half the picture. The second invention turns the AI's attention to the ground the car is actually driving on, which is rarely the perfectly flat plane that simple systems assume.
The model treats the ground as a lattice of points with known X and Y coordinates and predicts the missing piece, the Z value, which is the elevation of each point. Connect those points and you get a 3D mesh of the road itself, revealing hillcrests, downhill grades, banked ramps, speed bumps, and potholes.
Every point carries attributes too. The model can tell whether a surface is drivable, what material it is (grass, dirt, asphalt, or concrete), whether it's road or sidewalk, whether it has paint and what color, and where the curbs are. Picture the car labeling every patch of ground in view with something like "asphalt, one meter above us, drivable, yellow paint."
Training is clever here as well. A 3D point cloud of the surroundings, from LiDAR or triangulated from camera images, gets paired with a segmentation network that labels pixels as road, curb, sidewalk, and so on. Points belonging to the same surface are grouped, a smooth mesh is fitted over them to wipe out noise, and that clean mesh becomes the answer key.
One subtle fix really stands out. When your car rolls over a speed bump, everything measured relative to the car seems to jump, even though the world didn't move at all. So the coordinates are anchored to the road surface instead of the car, which smooths away the jitter caused by the car's own bouncing.
The real magic appears when this invention teams up with the first one. In Tesla's example, occupancy detection covers a vertical window of about 3 meters above and below the car, so a traffic cone sitting on a steep downhill road 4 meters below you would fall outside it.
The surface model solves this by first figuring out that the road there sits 4 meters lower. The cone is then measured against the hill's surface rather than against the car, and the two results are combined. Separating "where is the ground" from "what's on the ground" effectively stretches how far the car can see.
Because all of this is computed from images on the fly, the car can understand a road it has never driven before without first pinpointing itself on a prebuilt map.
🧠 JOB THREE: DECIDING WHO GOES FIRST
A perfect 3D picture means nothing if the car can't decide what to do with it. Tesla calls the third invention joint behavior planning and forecasting, and that name gives the game away.
Tesla's example is a classic: you're turning left into the far lane, a pedestrian is crossing that lane, and another car is turning right onto the same road behind them. It's the same kind of unprotected left that Tesla walked through on stage at AI Day 2022.
The traditional approach would forecast every possible path for the pedestrian and the other car, then grind through an optimization over all of your own possible paths. Tesla points out how much processing power, time, and energy that burns, especially when decisions may be made as often as every 50 milliseconds, or 20 times per second.
Tesla's answer is a hierarchical nodal graph, which works a lot like a choose-your-own-adventure book. At the top sits a goal layer holding what the car wants to achieve, like turning left into the far lane. Below it come layers of interaction nodes, each one a different way the encounter could play out.
The first branch might be "drive in front of the pedestrian" versus "yield and turn behind them." Under the yield branch, the next layer splits again into "slip in ahead of the other car" or "let both pass first." Each node bundles your move and the other road users' moves into one package, so planning and forecasting merge into a single, much simpler question: in what order do we pass each other?
Every node gets a score, which can come from a neural network. The score weighs four things: physics checks such as collisions, ride comfort, how likely a human would be to take over, and a human-like discriminator that judges whether a human driver would make the same choice.
A path's total score combines the scores of its nodes, and the highest-scoring path wins and steers the car.
The efficiency tricks are where this really shines. A node is only expanded when its score clears a threshold, so a branch like "cut in front of the pedestrian" can be dropped on the spot and never explored further. When the graph hits a size cap, the weakest node gets cut, keeping the workload steady instead of spiking when a crowd suddenly shows up.
The fleet feeds this graph too. A server can merge scenarios reported by many cars into one master graph, skip duplicate nodes, ship it to the vehicles, and push updated versions over time. And once the turn is complete, a fresh goal like "drive straight in the new lane" kicks off the next round.
🅿️ JOB FOUR: PARKING WITH INCH-LEVEL PRECISION
The fourth invention takes everything into the tightest spaces of all. A voxel grid that works great on the open road isn't precise enough when you're squeezing into a parking garage, so this one cranks occupancy up to high fidelity.
The key concept is the signed distance field, or SDF. Instead of a simple occupied-or-empty label, each voxel stores a precise number: its distance to the nearest surface. Positive values sit outside an object, negative values sit inside it, and zero lands exactly on the surface, much like contour lines on a hiking map trace the true shape of a mountain instead of just marking where it is.
To turn those numbers into a picture, the rendering works like a 3D printer building layer by layer. It creates a thin horizontal slice, about an inch thick in Tesla's example, starting at road level, coloring the voxels with negative values and leaving the positive ones transparent. Then it builds the next slice up, and the next, and stacks them all into a complete 3D scene.
Tesla shows the difference side by side. A raw point cloud is memory-hungry and noisy, a plain voxel grid looks blocky, and the SDF version captures the fine shapes needed to thread through tight spaces. A pole is drawn from the voxels the model actually sees, rather than by recognizing "pole" and dropping a stock 3D model into the scene.
Since an SDF measures distance to the nearest surface rather than to your car, the system casts virtual rays from camera pixels into the grid to pin down exactly how far away things are. Nearby objects can then be colored red and medium-range ones yellow.
Paint gets the same voxel-level treatment. Instead of hunting for continuous lines, the model checks whether each voxel of ground carries paint, which lets crosswalks and accessible-parking symbols come through along with their color and thickness.
If you've used High Fidelity Park Assist, which Tesla describes as a 3D view showing the shape and proximity of nearby objects, barriers, vehicles, and painted road markings, this will feel very familiar.
Then the parking flow kicks in:
🔍 The car recognizes it has entered a parking area using cues like its slow speed, the map, a sign listing parking prices, or how nearby cars are positioned.
🧱 The high-fidelity model reconstructs the space, and a model trained on labeled parking lots tells real parking spots apart from random empty space.
🚫 Spots get ruled out by "no parking" signs, time-limited signs checked against the current time, red-painted curbs, or an accessible-parking symbol when the driver isn't entitled to use it.
🏆 The survivors get a suitability score that favors spots closer to you and requiring fewer turns and fewer switches between drive and reverse, and you can even ask to park away from other cars.
👆 The top spot gets a "P" marker, you tap the one you want, confirm, and the car parks itself while the screen tracks its progress.
Scores keep updating as you drive, so a spot that looked awkward can jump to the top after you make a turn. Because paint alone can define a spot, it works in a wide-open stadium lot with no curbs in sight, where Tesla notes that ultrasonic-only systems fall short because they can't see paint at all.
It handles both perpendicular and parallel spots, and anyone who has used tap-to-park Autopark will recognize that "P" icon instantly.
🧩 HOW THE PIECES LOCK TOGETHER
Lay all four side by side and the architecture is gorgeous. Occupancy says what's solid and what's moving, the surface model describes the ground under everything, the planner turns it all into a decision, and the high-fidelity layer adds inch-level precision for tight spaces. Tesla even names occupancy as a way for the planner to detect the road users it reasons about, so the links are built right in.
They share the same DNA as well. Cameras go in, 3D understanding comes out, and data from the fleet keeps flowing back to improve the models.
The sharpest owners in the replies will raise a fair point here: FSD went end-to-end with v12, replacing much of the hand-written planning code with a single neural network. That's true, and three of these four inventions date back to 2022, well before that switch.
My take is that the ideas didn't disappear; they got absorbed. Ashok Elluswamy, Tesla's AI software chief and a named inventor on three of the four, has described today's FSD as one network running from camera video straight to control, while noting it can still be probed for its 3D understanding. And the camera-based parking visuals owners use today line up closely with what the fourth invention describes.
🧠 ONE BRAIN, MANY BODIES
Here's where it gets truly exciting. Look closely, and none of these four jobs is really about cars. Seeing what's there, reading the surface underfoot, deciding who goes first, and settling precisely into a spot are the same four jobs facing any machine that moves through our world.
On the road, that's exactly the skill set a Robotaxi needs when nobody is sitting in the driver's seat. But Tesla is clearly thinking bigger. Right alongside cars and trucks, it describes a general-purpose bipedal humanoid robot as one of the machines that can run this technology. Tesla's AI software chief has said the same foundational model behind FSD is being adapted for Optimus.
Picture what that means. For a car, reading the ground means spotting a curb or a hill. For a humanoid robot, the same skill could mean handling a staircase or a ramp. Deciding who goes first at an intersection becomes deciding who goes first in a narrow hallway. Parking becomes walking back to a charging station and docking, which Tesla already sketches out for a robot moving through an office full of desks, chairs, and people.
Your Tesla may be the first body this camera-only brain learned to control. It's very unlikely to be the last.
📸 FROM FLAT PIXELS TO A 3D WORLD
Every camera on your Tesla sees the world as a flat picture. Your car, though, has d that is anything but flat, packed with parked cars, sloping roads, painted lines, and pedestrians about to step off the curb. Turning those flat pictures into confident moves, using cameras alone with no LiDAR or radar on the road, is one of the hardest problems in autonomy.
Tesla has been designing its answer piece by piece, and four of its inventions, read together, tell the whole story. They trace the journey from raw camera pixels to a car that understands its surroundings, decides what to do, and parks itself in the exact spot you tapped.
Think about everything you do without even noticing when you drive. You spot the car stopped ahead, feel the road dip into a hill, decide whether to let the pedestrian cross first, and then squeeze into a tight spot without kissing the pillar. Those four instincts are exactly what Tesla is teaching its cameras.
And here's the beautiful part: all four run on the same engine. At the heart of each one sits a single pipeline that turns your car's camera feeds into a living 3D model of the world around it. Each invention then teaches that engine a brand-new skill, so let's break them down one by one.
🧊 JOB ONE: SEEING WHAT'S THERE
The first invention, already officially granted, tackles the most basic question of all. It slices the space around the car into voxels, which are simply 3D pixels, like a giant invisible Minecraft grid wrapped around the vehicle. For every voxel, the AI predicts whether that little cube of space is occupied by something with mass.
The genius here is that the system doesn't need to know what the object is. A car, a pedestrian, or a traffic sign in your path all get the same treatment, because a solid voxel in front of you is something to avoid no matter what it's called. That makes the approach naturally robust against strange, one-off objects that no classifier was ever trained to name.
The pipeline is a work of art. Images from each camera are rectified and compressed into features by image networks (RegNet and BiFPN in Tesla's example). A transformer then lets every 3D location "ask" the camera images which pixels belong to it, fusing all the separate views into one unified 3D space.
Next, the current 3D snapshot is stacked with the previous few, aligned using the car's own trajectory, so the model sees both space and time. Deconvolution layers then expand those fused features back into a full voxel grid.
The outputs go far beyond a simple yes or no. The model can predict occupancy flow, meaning how fast each occupied voxel is moving, plus 3D semantics, such as whether a cluster of voxels is a curb, a building, or a moving vehicle.
The resolution is smart too. Tesla's example uses 33 cm voxels by default but shrinks them to around 10 cm when they're occupied and close to the car or the driving surface. If a curb fills only 16 cm of a 33 cm voxel, a regression step estimates how much of it is actually solid, and interpolation lets software query any point inside a voxel.
So how does a camera-only model learn 3D in the first place? During training, Tesla can use sensor data such as LiDAR to build the answer key, then teach the model to reproduce it from camera footage. Once trained, Tesla is crystal clear on this point: the model predicts 3D occupancy using only the 2D images from the cameras.
The final result is a queryable dataset. Any other piece of software can simply ask "is the point at X, Y, Z occupied?" and get an instant answer.
🛣️ JOB TWO: READING THE ROAD
Knowing where the obstacles are is only half the picture. The second invention turns the AI's attention to the ground the car is actually driving on, which is rarely the perfectly flat plane that simple systems assume.
The model treats the ground as a lattice of points with known X and Y coordinates and predicts the missing piece, the Z value, which is the elevation of each point. Connect those points and you get a 3D mesh of the road itself, revealing hillcrests, downhill grades, banked ramps, speed bumps, and potholes.
Every point carries attributes too. The model can tell whether a surface is drivable, what material it is (grass, dirt, asphalt, or concrete), whether it's road or sidewalk, whether it has paint and what color, and where the curbs are. Picture the car labeling every patch of ground in view with something like "asphalt, one meter above us, drivable, yellow paint."
Training is clever here as well. A 3D point cloud of the surroundings, from LiDAR or triangulated from camera images, gets paired with a segmentation network that labels pixels as road, curb, sidewalk, and so on. Points belonging to the same surface are grouped, a smooth mesh is fitted over them to wipe out noise, and that clean mesh becomes the answer key.
One subtle fix really stands out. When your car rolls over a speed bump, everything measured relative to the car seems to jump, even though the world didn't move at all. So the coordinates are anchored to the road surface instead of the car, which smooths away the jitter caused by the car's own bouncing.
The real magic appears when this invention teams up with the first one. In Tesla's example, occupancy detection covers a vertical window of about 3 meters above and below the car, so a traffic cone sitting on a steep downhill road 4 meters below you would fall outside it.
The surface model solves this by first figuring out that the road there sits 4 meters lower. The cone is then measured against the hill's surface rather than against the car, and the two results are combined. Separating "where is the ground" from "what's on the ground" effectively stretches how far the car can see.
Because all of this is computed from images on the fly, the car can understand a road it has never driven before without first pinpointing itself on a prebuilt map.
🧠 JOB THREE: DECIDING WHO GOES FIRST
A perfect 3D picture means nothing if the car can't decide what to do with it. Tesla calls the third invention joint behavior planning and forecasting, and that name gives the game away.
Tesla's example is a classic: you're turning left into the far lane, a pedestrian is crossing that lane, and another car is turning right onto the same road behind them. It's the same kind of unprotected left that Tesla walked through on stage at AI Day 2022.
The traditional approach would forecast every possible path for the pedestrian and the other car, then grind through an optimization over all of your own possible paths. Tesla points out how much processing power, time, and energy that burns, especially when decisions may be made as often as every 50 milliseconds, or 20 times per second.
Tesla's answer is a hierarchical nodal graph, which works a lot like a choose-your-own-adventure book. At the top sits a goal layer holding what the car wants to achieve, like turning left into the far lane. Below it come layers of interaction nodes, each one a different way the encounter could play out.
The first branch might be "drive in front of the pedestrian" versus "yield and turn behind them." Under the yield branch, the next layer splits again into "slip in ahead of the other car" or "let both pass first." Each node bundles your move and the other road users' moves into one package, so planning and forecasting merge into a single, much simpler question: in what order do we pass each other?
Every node gets a score, which can come from a neural network. The score weighs four things: physics checks such as collisions, ride comfort, how likely a human would be to take over, and a human-like discriminator that judges whether a human driver would make the same choice.
A path's total score combines the scores of its nodes, and the highest-scoring path wins and steers the car.
The efficiency tricks are where this really shines. A node is only expanded when its score clears a threshold, so a branch like "cut in front of the pedestrian" can be dropped on the spot and never explored further. When the graph hits a size cap, the weakest node gets cut, keeping the workload steady instead of spiking when a crowd suddenly shows up.
The fleet feeds this graph too. A server can merge scenarios reported by many cars into one master graph, skip duplicate nodes, ship it to the vehicles, and push updated versions over time. And once the turn is complete, a fresh goal like "drive straight in the new lane" kicks off the next round.
🅿️ JOB FOUR: PARKING WITH INCH-LEVEL PRECISION
The fourth invention takes everything into the tightest spaces of all. A voxel grid that works great on the open road isn't precise enough when you're squeezing into a parking garage, so this one cranks occupancy up to high fidelity.
The key concept is the signed distance field, or SDF. Instead of a simple occupied-or-empty label, each voxel stores a precise number: its distance to the nearest surface. Positive values sit outside an object, negative values sit inside it, and zero lands exactly on the surface, much like contour lines on a hiking map trace the true shape of a mountain instead of just marking where it is.
To turn those numbers into a picture, the rendering works like a 3D printer building layer by layer. It creates a thin horizontal slice, about an inch thick in Tesla's example, starting at road level, coloring the voxels with negative values and leaving the positive ones transparent. Then it builds the next slice up, and the next, and stacks them all into a complete 3D scene.
Tesla shows the difference side by side. A raw point cloud is memory-hungry and noisy, a plain voxel grid looks blocky, and the SDF version captures the fine shapes needed to thread through tight spaces. A pole is drawn from the voxels the model actually sees, rather than by recognizing "pole" and dropping a stock 3D model into the scene.
Since an SDF measures distance to the nearest surface rather than to your car, the system casts virtual rays from camera pixels into the grid to pin down exactly how far away things are. Nearby objects can then be colored red and medium-range ones yellow.
Paint gets the same voxel-level treatment. Instead of hunting for continuous lines, the model checks whether each voxel of ground carries paint, which lets crosswalks and accessible-parking symbols come through along with their color and thickness.
If you've used High Fidelity Park Assist, which Tesla describes as a 3D view showing the shape and proximity of nearby objects, barriers, vehicles, and painted road markings, this will feel very familiar.
Then the parking flow kicks in:
🔍 The car recognizes it has entered a parking area using cues like its slow speed, the map, a sign listing parking prices, or how nearby cars are positioned.
🧱 The high-fidelity model reconstructs the space, and a model trained on labeled parking lots tells real parking spots apart from random empty space.
🚫 Spots get ruled out by "no parking" signs, time-limited signs checked against the current time, red-painted curbs, or an accessible-parking symbol when the driver isn't entitled to use it.
🏆 The survivors get a suitability score that favors spots closer to you and requiring fewer turns and fewer switches between drive and reverse, and you can even ask to park away from other cars.
👆 The top spot gets a "P" marker, you tap the one you want, confirm, and the car parks itself while the screen tracks its progress.
Scores keep updating as you drive, so a spot that looked awkward can jump to the top after you make a turn. Because paint alone can define a spot, it works in a wide-open stadium lot with no curbs in sight, where Tesla notes that ultrasonic-only systems fall short because they can't see paint at all.
It handles both perpendicular and parallel spots, and anyone who has used tap-to-park Autopark will recognize that "P" icon instantly.
🧩 HOW THE PIECES LOCK TOGETHER
Lay all four side by side and the architecture is gorgeous. Occupancy says what's solid and what's moving, the surface model describes the ground under everything, the planner turns it all into a decision, and the high-fidelity layer adds inch-level precision for tight spaces. Tesla even names occupancy as a way for the planner to detect the road users it reasons about, so the links are built right in.
They share the same DNA as well. Cameras go in, 3D understanding comes out, and data from the fleet keeps flowing back to improve the models.
The sharpest owners in the replies will raise a fair point here: FSD went end-to-end with v12, replacing much of the hand-written planning code with a single neural network. That's true, and three of these four inventions date back to 2022, well before that switch.
My take is that the ideas didn't disappear; they got absorbed. Ashok Elluswamy, Tesla's AI software chief and a named inventor on three of the four, has described today's FSD as one network running from camera video straight to control, while noting it can still be probed for its 3D understanding. And the camera-based parking visuals owners use today line up closely with what the fourth invention describes.
🧠 ONE BRAIN, MANY BODIES
Here's where it gets truly exciting. Look closely, and none of these four jobs is really about cars. Seeing what's there, reading the surface underfoot, deciding who goes first, and settling precisely into a spot are the same four jobs facing any machine that moves through our world.
On the road, that's exactly the skill set a Robotaxi needs when nobody is sitting in the driver's seat. But Tesla is clearly thinking bigger. Right alongside cars and trucks, it describes a general-purpose bipedal humanoid robot as one of the machines that can run this technology. Tesla's AI software chief has said the same foundational model behind FSD is being adapted for Optimus.
Picture what that means. For a car, reading the ground means spotting a curb or a hill. For a humanoid robot, the same skill could mean handling a staircase or a ramp. Deciding who goes first at an intersection becomes deciding who goes first in a narrow hallway. Parking becomes walking back to a charging station and docking, which Tesla already sketches out for a robot moving through an office full of desks, chairs, and people.
Your Tesla may be the first body this camera-only brain learned to control. It's very unlikely to be the last.
5
4
31
4,690
📚 SOURCES
✅ Job one: US 12,469,160 B2, Artificial intelligence modeling techniques for vision-based occupancy determination (granted Nov 11, 2025)
✅ Job two: US 2026/0330841 A1, Artificial intelligence modeling techniques for vision-based surface determination (published Oct 22, 2026)
✅ Job three: US 2026/0105614 A1, Artificial intelligence modeling techniques for joint behavior planning and forecasting (published Apr 16, 2026)
✅ Job four: WO 2025/193615 A1, Artificial intelligence modeling techniques for vision-based high-fidelity occupancy determination and assisted parking applications (published Sep 18, 2025)
✅ Job one: US 12,469,160 B2, Artificial intelligence modeling techniques for vision-based occupancy determination (granted Nov 11, 2025)
✅ Job two: US 2026/0330841 A1, Artificial intelligence modeling techniques for vision-based surface determination (published Oct 22, 2026)
✅ Job three: US 2026/0105614 A1, Artificial intelligence modeling techniques for joint behavior planning and forecasting (published Apr 16, 2026)
✅ Job four: WO 2025/193615 A1, Artificial intelligence modeling techniques for vision-based high-fidelity occupancy determination and assisted parking applications (published Sep 18, 2025)
0
0
3
973
GOOD NEWS 🚖 According to the latest Texas DMV filings, Tesla has registered an additional 11 Cybercabs, pushing the state's total registered Cybercab count to 431 vehicles, accounting for nearly 46% of the state's entire autonomous fleet 🔥
With 507 commercial Model Ys ide them, Tesla's total commercial Robotaxi fleet in Texas has expanded to 938 vehicles 🔥
The total fleet rapidly closes in on the 1,000-vehicle mark 🔥
With 507 commercial Model Ys ide them, Tesla's total commercial Robotaxi fleet in Texas has expanded to 938 vehicles 🔥
The total fleet rapidly closes in on the 1,000-vehicle mark 🔥
6
6
90
10,229
PURE VISION BREAKTHROUGH 🚨 Tesla has designed a brilliant sixth sense: cameras that can feel the shape of the road 👁️
For one heartbeat at the top of a steep hill, the windshield shows nothing but sky. Your brain still knows the road is about to plunge, and a car driving on ow it too.
Tesla's cameras are already sharp at spotting obstacles. The occupancy network chops the world around the car into voxels, 3D cubes that may default to about 33 cm, and flags every cube holding something with mass, whether it's moving, and what it probably is.
The road itself is a different beast. It isn't an obstacle at all. It's the stage everything else stands on, and the car needs to know its height, its edges, its material, and its paint.
A flat top-down view crushes the hills out of the picture. An occupancy check may only reach a slice of height around the car, such as 3 meters up and down, and a steep downhill can drop right out of that window.
Conventional systems often lean on localization, pinning the car to a known map before trusting the road's shape. Tesla is chasing something bolder, a car that reads a road it has never seen with nothing but its own eyes.
So Tesla has designed a second job for its camera network, and it's a big one. On top of spotting obstacles, the network sculpts the ground itself, live, frame after frame.
It all starts with a familiar engine. Tesla rectifies the feeds from the car's cameras, eight in one example layout, and crunches them into features with a RegNet and BiFPN stack. A transformer then fires 3D queries into those 2D features and pulls everything into one shared 3D space around the car.
Next, the network gets a memory. Tesla aligns each fresh 3D snapshot using the car's own trajectory and fuses it with the last few moments, so the scene carries both space and time. Deconvolution layers then blow it back up into fine 3D detail.
Now comes the magic. Tesla lays a flat lattice of points across the ground, each with a known X and Y, and the model predicts the one number cameras never measure directly. That number is Z, the height of every point above a zero-elevation surface.
Every point also carries a feature vector, a tiny ID card for the ground. It can mark a spot as navigable or off-limits, as grass, dirt, asphalt, or concrete, as road or sidewalk, and it can even flag paint and its color.
Stitch the points together and the street snaps into a mesh, a stretchy 3D sheet draped over reality. Curbs pop out as their own surfaces, the drivable lanes split off cleanly, and a road plunging downhill ahead looks unmistakably different from one climbing uphill off to the side.
Building the answer key is a feat of its own. Data-collection cars may capture a point cloud of the scene with sensors like LiDAR, or Tesla can triangulate the points straight from camera images. Every 3D point then gets projected into the camera frames from the same moments.
From there, a segmentation network labels those pixels as tree, sky, curb, or road, plus material and paint. Tesla sorts the points by category, drapes a smooth mesh over each group to scrub out noise, and that clean surface becomes the target the main model chases.
That homework pays off big. Once trained, the model may need nothing but images, because it has learned how a painted uphill dirt road or a concrete curb looks through a lens. Even on a street it has never seen, it can rebuild the surface on the spot.
The trick isn't locked to cars, either. A truck or a bipedal humanoid robot could carry the same camera-built ground model to read the terrain under its wheels or feet.
Tesla even tames the bounce. When the car hops a speed bump or dips into a pothole, coordinates measured from the car jump around. Tesla anchors the heights to the road surface instead, so the jolt gets smoothed out rather than shaking the whole map.
Then the surface map teams up with occupancy, and the pair gets downright crafty. When a traffic cone sits on a hill ahead, the model can work out the hill first, size up the cone as if it stood on flat ground, and then stack the cone right back on top of the hill.
That split stretches the car's reach past the edge of its occupancy window. A cone on a downhill 4 meters below the car sits outside a ±3 meter range, yet with the surface model supplying the hill's height, it can still land in the right spot.
All of it pours into a queryable dataset. Driving software can ping any spot around the car for its height and attributes, and the screen can paint lanes and curbs in their own colors. A neural radiance field may even render a lifelike 3D twin of the surroundings.
Back on that hilltop, the windshield still shows nothing but sky. Behind it, the car's cameras have already sculpted the street in 3D. They can tell how steep the crest is and which way the road is about to drop.
That's the sixth sense. Tesla's cameras already see what sits on the road. With this design, they can feel the road itself.
For one heartbeat at the top of a steep hill, the windshield shows nothing but sky. Your brain still knows the road is about to plunge, and a car driving on ow it too.
Tesla's cameras are already sharp at spotting obstacles. The occupancy network chops the world around the car into voxels, 3D cubes that may default to about 33 cm, and flags every cube holding something with mass, whether it's moving, and what it probably is.
The road itself is a different beast. It isn't an obstacle at all. It's the stage everything else stands on, and the car needs to know its height, its edges, its material, and its paint.
A flat top-down view crushes the hills out of the picture. An occupancy check may only reach a slice of height around the car, such as 3 meters up and down, and a steep downhill can drop right out of that window.
Conventional systems often lean on localization, pinning the car to a known map before trusting the road's shape. Tesla is chasing something bolder, a car that reads a road it has never seen with nothing but its own eyes.
So Tesla has designed a second job for its camera network, and it's a big one. On top of spotting obstacles, the network sculpts the ground itself, live, frame after frame.
It all starts with a familiar engine. Tesla rectifies the feeds from the car's cameras, eight in one example layout, and crunches them into features with a RegNet and BiFPN stack. A transformer then fires 3D queries into those 2D features and pulls everything into one shared 3D space around the car.
Next, the network gets a memory. Tesla aligns each fresh 3D snapshot using the car's own trajectory and fuses it with the last few moments, so the scene carries both space and time. Deconvolution layers then blow it back up into fine 3D detail.
Now comes the magic. Tesla lays a flat lattice of points across the ground, each with a known X and Y, and the model predicts the one number cameras never measure directly. That number is Z, the height of every point above a zero-elevation surface.
Every point also carries a feature vector, a tiny ID card for the ground. It can mark a spot as navigable or off-limits, as grass, dirt, asphalt, or concrete, as road or sidewalk, and it can even flag paint and its color.
Stitch the points together and the street snaps into a mesh, a stretchy 3D sheet draped over reality. Curbs pop out as their own surfaces, the drivable lanes split off cleanly, and a road plunging downhill ahead looks unmistakably different from one climbing uphill off to the side.
Building the answer key is a feat of its own. Data-collection cars may capture a point cloud of the scene with sensors like LiDAR, or Tesla can triangulate the points straight from camera images. Every 3D point then gets projected into the camera frames from the same moments.
From there, a segmentation network labels those pixels as tree, sky, curb, or road, plus material and paint. Tesla sorts the points by category, drapes a smooth mesh over each group to scrub out noise, and that clean surface becomes the target the main model chases.
That homework pays off big. Once trained, the model may need nothing but images, because it has learned how a painted uphill dirt road or a concrete curb looks through a lens. Even on a street it has never seen, it can rebuild the surface on the spot.
The trick isn't locked to cars, either. A truck or a bipedal humanoid robot could carry the same camera-built ground model to read the terrain under its wheels or feet.
Tesla even tames the bounce. When the car hops a speed bump or dips into a pothole, coordinates measured from the car jump around. Tesla anchors the heights to the road surface instead, so the jolt gets smoothed out rather than shaking the whole map.
Then the surface map teams up with occupancy, and the pair gets downright crafty. When a traffic cone sits on a hill ahead, the model can work out the hill first, size up the cone as if it stood on flat ground, and then stack the cone right back on top of the hill.
That split stretches the car's reach past the edge of its occupancy window. A cone on a downhill 4 meters below the car sits outside a ±3 meter range, yet with the surface model supplying the hill's height, it can still land in the right spot.
All of it pours into a queryable dataset. Driving software can ping any spot around the car for its height and attributes, and the screen can paint lanes and curbs in their own colors. A neural radiance field may even render a lifelike 3D twin of the surroundings.
Back on that hilltop, the windshield still shows nothing but sky. Behind it, the car's cameras have already sculpted the street in 3D. They can tell how steep the crest is and which way the road is about to drop.
That's the sixth sense. Tesla's cameras already see what sits on the road. With this design, they can feel the road itself.
25
43
292
14,323
BOOOM 🇨🇳 When it comes to generating hype, nobody does it quite like Tesla 🤩
Over at Wanda Plaza in Yulin, China, the grand opening of their newest showroom became an absolute showstopper—and yes, they even brought out the Cybertruck, drawing a massive crowd of stunned right next door, several other auto brands running outdoor displays were left in total ghost-town mode, completely ignored 😱
You really have to wonder what was going through those sales reps' heads 😂
Over at Wanda Plaza in Yulin, China, the grand opening of their newest showroom became an absolute showstopper—and yes, they even brought out the Cybertruck, drawing a massive crowd of stunned right next door, several other auto brands running outdoor displays were left in total ghost-town mode, completely ignored 😱
You really have to wonder what was going through those sales reps' heads 😂
6
12
176
9,577
SOS 🇯🇵 Starting January 2027, Japan's national EV subsidy for Tesla is dropping 40%—from ¥1.27 million down to ¥762,000 😱
That is an immediate ¥508,000 (~$3,250) price jump for buyers:
⚠️ Model 3: Starting at ¥4.55M, up from ¥4.04M
⚠️ Model Y: Starting at ¥4.83M, up from ¥4.32M
lock in the 2026 rate, but deliveries must be registered by December 31 before the subsidy budget runs out 🐰
That is an immediate ¥508,000 (~$3,250) price jump for buyers:
⚠️ Model 3: Starting at ¥4.55M, up from ¥4.04M
⚠️ Model Y: Starting at ¥4.83M, up from ¥4.32M
lock in the 2026 rate, but deliveries must be registered by December 31 before the subsidy budget runs out 🐰
6
9
35
5,225
ENGINEERING NEWS 📐 Tesla has revived its daring Cybertruck exoskeleton: one metal sheet that carries the load 😇
On November 21, 2019, the day Tesla drove the Cybertruck onstage, the company also filed the first patent application for the truck's boldest promise, an took a more practical path, but a new application from that same family, published this week, reveals just how radical the original idea was.
On a typical car door, the shiny outer panel is mostly a cover. The real strength hides behind it, in a stack of parts most drivers never see.
Conventional doors are generally built in layers. Several stamped reinforcements get welded to an inner panel, and that whole inner assembly is then hemmed to the outer skin around the edges.
Buried in that sandwich is an anti-intrusion bar, usually a thick beam of steel or aluminum that helps shield passengers when another car slams into the door. Behind it, a mid-panel shoulders the load of other door components.
That bar exists because the outer panel, on its own, just doesn't have much strength. Every extra layer piles on cost and assembly time, and it makes everything inside the door harder to install and service.
Even today's Cybertruck follows a version of this layered recipe. Teardowns show its doors wrapping stainless inner and outer panels around a sealed module that holds the window regulator, while the stainless skin is glued onto a high-strength stamped steel structure.
Tesla's answer is to promote the outer panel from cover to skeleton. The door's exterior becomes one monolithic sheet of metal with a continuous, uninterrupted surface, and that single sheet forms a structural piece of the vehicle's exoskeleton.
The company can start with a rolled metal sheet such as cold-rolled steel, cut it to shape (laser cutting is one option), and form it into a door. The drawings show the result on a sharp-edged pickup that any Cybertruck owner will recognize in a heartbeat. And if the metal resists corrosion well enough on its own, the door may skip paint entirely, a look those owners already know well.
The clever part is where everything else attaches. A hinge fixes the sheet directly to the vehicle frame. The electric window assembly, meaning the glass and the motor that drives it, mounts directly onto that same sheet instead of hanging off a separate mid-panel.
That creates a clean structural load path. The window's weight and every force its motor generates flow from the window assembly into the outer panel, through the hinge, and straight into the frame.
In other words, the skin stops being decoration. It becomes the primary structural support for the window assembly, a job that used to belong to those hidden inner layers.
To pull that off, the panel has to take everything a door goes through over its lifetime. It handles the loads of opening and closing, and it can be built to shrug off abuse like being yanked open too far, slammed shut, or twisted, plus the push of the door seals and the tug of the window running up and down.
Because the sheet is made from tough metal like steel or aluminum, the separate anti-intrusion bar and mid-panel can go. The outer panel itself can deliver side impact protection as a primary structural member, while an interior trim panel still hides the window motor and other hardware on the cabin side.
The same thinking reaches the pillar between the front and rear doors. A conventional pillar backs its outer frame section with an inner support plus a second intermediate support. The exoskeleton version needs just one inner support, because the outer section is already strong enough to protect occupants in a side impact.
Every reinforcement that disappears is one fewer part to weld. With fewer of them to join, Tesla can cut down the assembly lines needed before the body comes together, which may significantly shrink both the manufacturing footprint and the cost.
And the idea doesn't stop at doors. Roof, hood, fender, trunk, and liftgate panels can all work the same way, with hinges, latches, and bolts mounted straight onto the metal, so panels that used to just cover the vehicle start pulling structural weight.
The Cybertruck rolled onstage in 2019 promising an exoskeleton, and the production truck still keeps much of its strength tucked behind the skin. This design turns the door inside out. The steel you run your hand across becomes the very steel holding the whole door together.
On November 21, 2019, the day Tesla drove the Cybertruck onstage, the company also filed the first patent application for the truck's boldest promise, an took a more practical path, but a new application from that same family, published this week, reveals just how radical the original idea was.
On a typical car door, the shiny outer panel is mostly a cover. The real strength hides behind it, in a stack of parts most drivers never see.
Conventional doors are generally built in layers. Several stamped reinforcements get welded to an inner panel, and that whole inner assembly is then hemmed to the outer skin around the edges.
Buried in that sandwich is an anti-intrusion bar, usually a thick beam of steel or aluminum that helps shield passengers when another car slams into the door. Behind it, a mid-panel shoulders the load of other door components.
That bar exists because the outer panel, on its own, just doesn't have much strength. Every extra layer piles on cost and assembly time, and it makes everything inside the door harder to install and service.
Even today's Cybertruck follows a version of this layered recipe. Teardowns show its doors wrapping stainless inner and outer panels around a sealed module that holds the window regulator, while the stainless skin is glued onto a high-strength stamped steel structure.
Tesla's answer is to promote the outer panel from cover to skeleton. The door's exterior becomes one monolithic sheet of metal with a continuous, uninterrupted surface, and that single sheet forms a structural piece of the vehicle's exoskeleton.
The company can start with a rolled metal sheet such as cold-rolled steel, cut it to shape (laser cutting is one option), and form it into a door. The drawings show the result on a sharp-edged pickup that any Cybertruck owner will recognize in a heartbeat. And if the metal resists corrosion well enough on its own, the door may skip paint entirely, a look those owners already know well.
The clever part is where everything else attaches. A hinge fixes the sheet directly to the vehicle frame. The electric window assembly, meaning the glass and the motor that drives it, mounts directly onto that same sheet instead of hanging off a separate mid-panel.
That creates a clean structural load path. The window's weight and every force its motor generates flow from the window assembly into the outer panel, through the hinge, and straight into the frame.
In other words, the skin stops being decoration. It becomes the primary structural support for the window assembly, a job that used to belong to those hidden inner layers.
To pull that off, the panel has to take everything a door goes through over its lifetime. It handles the loads of opening and closing, and it can be built to shrug off abuse like being yanked open too far, slammed shut, or twisted, plus the push of the door seals and the tug of the window running up and down.
Because the sheet is made from tough metal like steel or aluminum, the separate anti-intrusion bar and mid-panel can go. The outer panel itself can deliver side impact protection as a primary structural member, while an interior trim panel still hides the window motor and other hardware on the cabin side.
The same thinking reaches the pillar between the front and rear doors. A conventional pillar backs its outer frame section with an inner support plus a second intermediate support. The exoskeleton version needs just one inner support, because the outer section is already strong enough to protect occupants in a side impact.
Every reinforcement that disappears is one fewer part to weld. With fewer of them to join, Tesla can cut down the assembly lines needed before the body comes together, which may significantly shrink both the manufacturing footprint and the cost.
And the idea doesn't stop at doors. Roof, hood, fender, trunk, and liftgate panels can all work the same way, with hinges, latches, and bolts mounted straight onto the metal, so panels that used to just cover the vehicle start pulling structural weight.
The Cybertruck rolled onstage in 2019 promising an exoskeleton, and the production truck still keeps much of its strength tucked behind the skin. This design turns the door inside out. The steel you run your hand across becomes the very steel holding the whole door together.
8
8
98
7,530
SOS 🚨 Fuel prices just rocketed 17 spots to become the #1 issue in trucking, according to ATRI's 2026 survey of more than 3,000 carriers, drivers and industry stakeholders 😳
It's only the fourth time in 22 years that fuel has topped the list, and the first time ever that ed it #1 😳
In California, diesel shot past $8 a gallon this year 😱
But diesel prices can't hurt a truck that doesn't burn diesel. That's the Tesla Semi's pitch 🔥
The Semi is rated at 1.7 kWh per mile at 82,000 lbs, and DHL clocked 1.72 kWh per mile on a fully loaded 390-mile run. Here's how the math shakes out:
⛽ Diesel truck at ~6.5 mpg and $8/gal: about $1.23 per mile
⚡ Tesla Semi at 1.7 kWh/mile and $0.30/kWh: about $0.51 per mile
Over 100,000 miles a year, that's more than $70,000 in fuel savings per truck.
Electricity also doesn't swing with the oil market the way diesel does. That turns the most unpredictable line in your operating budget into one you can actually plan around. Zero tailpipe emissions also take on #8 on the list, diesel emissions regulations, something California carriers know all too well.
Deliveries are rolling, the Nevada factory is ramping toward 50,000 trucks a year, and the future of freight is plugging in ⚡
It's only the fourth time in 22 years that fuel has topped the list, and the first time ever that ed it #1 😳
In California, diesel shot past $8 a gallon this year 😱
But diesel prices can't hurt a truck that doesn't burn diesel. That's the Tesla Semi's pitch 🔥
The Semi is rated at 1.7 kWh per mile at 82,000 lbs, and DHL clocked 1.72 kWh per mile on a fully loaded 390-mile run. Here's how the math shakes out:
⛽ Diesel truck at ~6.5 mpg and $8/gal: about $1.23 per mile
⚡ Tesla Semi at 1.7 kWh/mile and $0.30/kWh: about $0.51 per mile
Over 100,000 miles a year, that's more than $70,000 in fuel savings per truck.
Electricity also doesn't swing with the oil market the way diesel does. That turns the most unpredictable line in your operating budget into one you can actually plan around. Zero tailpipe emissions also take on #8 on the list, diesel emissions regulations, something California carriers know all too well.
Deliveries are rolling, the Nevada factory is ramping toward 50,000 trucks a year, and the future of freight is plugging in ⚡
4
5
21
2,660