Gemini 4 Argon taught a robot to play cornhole just by watching a video.
It mapped the physics, coded its own simulation, and taught @Apptronik's Apollo humanoid to sink 7 throws in a row.
It mapped the physics, coded its own simulation, and taught @Apptronik's Apollo humanoid to sink 7 throws in a row.
Oct 13, 2026 · 6:41 PM UTC
285
409
3,425
271,763
@Apptronik By turning video into a training ground for physical AI, Argon compresses weeks of manual modeling into just a few hours. ↓
goo.gle/3VxIB7v
goo.gle/3VxIB7v
17
7
113
27,615
@GoogleDeepMind @Apptronik The link between AI and real-world robotics keeps getting more interesting.
4
0
19
2,097
@GoogleDeepMind @Apptronik Watch video, map physics, generate simulation, train policy 🦾
0
0
3
201
@GoogleDeepMind @Apptronik I’m going to need the robot on my team at the next cookout.
0
0
2
213
@GoogleDeepMind i wonder if apollo used any misses to update the simulation between throws.
0
0
1
130
@GoogleDeepMind @Apptronik The cornhole is the fun part. The real story is Argon writing its own sim and doing system ID from one human video. If that generalizes, the slowest part of training humanoids just got a lot faster, for Apollo, Optimus, everybody.
0
0
1
181
@GoogleDeepMind @Apptronik Pretty cool. Especially the robots understanding the game. LLMs are still the king of generalization and reasoning tasks
0
0
1
206
@GoogleDeepMind @Apptronik learning physics from a single video is impressive, but cornhole is a closed system with fixed rules. the gap to generalize this to tasks where the rules aren't visible in one example is still wide.
0
0
0
146
@GoogleDeepMind @Apptronik A humanoid sinking seven cornhole throws in a row because Gemini watched a video and wrote its own physics sim is kinda nuts
What a time to be alive
What a time to be alive
0
0
20
1,346
@GoogleDeepMind @Apptronik The robots nailing cornhole before laundry tells you exactly who's funding them
2
0
7
742
@GoogleDeepMind @Apptronik Was Gemini 4 Argon also used to train the bots to reply and boost this post?
1
0
5
604
@GoogleDeepMind @Apptronik Next lesson: miss occasionally so people keep inviting you.
0
0
4
618
@GoogleDeepMind @Apptronik 7 in a row is cool, but the real number is how many throws it took to get there, and the post skips it.
3
0
5
631
@GoogleDeepMind @Apptronik Would be scary when it goes rouge and throws a grenade at you with that precision
1
1
2
322
@GoogleDeepMind @Apptronik 7 throws in a row 2 inches away from the hole 🤣
0
0
2
625
@GoogleDeepMind @Apptronik Writing its own sim from a video is the crazy part. We do the boring version on CNC: run the real firmware in sim first, and it catches stuff a G-code preview never shows. How close were the sim throws to the real ones?
0
0
2
157
@GoogleDeepMind @Apptronik How many tries did it need in the sim before it made real throws?
0
0
2
407
@GoogleDeepMind @Apptronik Gemini 4 Argon watched one cornhole video, figured out the physics, and taught a robot to sink seven in a row. The watching-and-learning bit is what used to take a robotics team forever. I still miss the bag on purpose
0
0
2
401
@GoogleDeepMind @Apptronik I'd put a different bag in its hand and move the board for the next run. Keep the misses in the video, too. That would tell us more about the physics it learned than another perfect streak.
0
0
2
126
@GoogleDeepMind @Apptronik First our jobs then our corn hole entertainment what is even left for the humans???
0
1
2
481
@GoogleDeepMind @Apptronik teaching it cornhole before laundry is a bold choice 😭
0
0
1
277
@GoogleDeepMind @Apptronik it wrote its own physics sim before throwing a single bag, which is more prep than any human has put into cornhole
3
0
1
448
@GoogleDeepMind @Apptronik 7 bags in a row after one video is the flex. My coding agent still needs two tries to cd into the right folder.
0
0
1
74
@GoogleDeepMind @Apptronik I’ve watched hours of cooking videos and still burn toast. This feels personal.
0
0
2
208
@GoogleDeepMind @Apptronik watched one video and sank 7 straight. i've played cornhole at every bbq for years and never done that lol
0
0
2
187
@GoogleDeepMind @Apptronik The 7-in-a-row is the flashy bit. What I want is whether Argon checked the sim against real throw physics before Apollo ran, or if it just optimized until the video looked right. Those are different products.
0
0
0
3
@GoogleDeepMind @Apptronik 7 in a row is better than my cornhole average. curious how many video demos it needed before the sim matched real throws
0
0
1
6
@GoogleDeepMind @Apptronik A self-built simulator could turn one demonstration into many practice attempts without using the robot.
0
0
1
195
@GoogleDeepMind @Apptronik Wow🎉 Very nice. But not sure about establishing physics simulation was the right strategy... My guess, inverse kinematics and fuzzy logic could be more organic perhaps faster way...
0
0
0
158
@GoogleDeepMind @Apptronik 27 feet...not 2 feet, ur boy needs to use chatGPT & check the rules.
0
0
0
216
@GoogleDeepMind @Apptronik 7 in a row is better than every uncle at every barbecue ever lol 😂😂😂😂😂
0
0
0
203
@GoogleDeepMind @Apptronik Crazy. Problem is robots will become as stupid as us or even more, because they will be better at stupidity as well than us.
0
0
0
24
@GoogleDeepMind @Apptronik Watch once, build its own sim, then practice. Pretty much how we learn too 😄 7 in a row is better than most of us at a backyard party
0
0
1
169
@GoogleDeepMind @Apptronik still far from an actual general rationalization…
0
0
0
91
