Pinned Tweet
Physical AI has a data problem. To operate in the real world, models need to learn from human experience.
The scale of what that demands is increasingly clear:
↳ $7T potential humanoid market by 2050
↳ $11.9B → $49.8B AI training dataset market by 2031
↳ ~100K hours of th tens of millions being targeted
Meeting that scale will require vastly more training data than exists today.
It’s where we’re increasingly focused at Poseidon.
The scale of what that demands is increasingly clear:
↳ $7T potential humanoid market by 2050
↳ $11.9B → $49.8B AI training dataset market by 2031
↳ ~100K hours of th tens of millions being targeted
Meeting that scale will require vastly more training data than exists today.
It’s where we’re increasingly focused at Poseidon.
45
9
114
37,467
Data in the Wild, ep. 004.
Task context
• Clean several pairs of shoes, then arrange by type + use frequency
• 11 minutes
• A home in India
Capture
• Handheld monocular Realme RMX3867
• 1280×720, H.264
Embodied behavior
• Repeated handling across shoes, cloth, cleaner + _shoe, rotate_shoe, apply_cleaner, change_pair, carry_shoes
• Challenges: object identity, state tracking, tool changes + long occlusions
Video sourced from Numo. Numo collects real-world data that Poseidon refines into training-ready datasets.
Long-horizon tasks require more than recognizing an action frame by frame. Models need to track which objects they’ve interacted with, how their state has changed, and what still needs to happen.
Task context
• Clean several pairs of shoes, then arrange by type + use frequency
• 11 minutes
• A home in India
Capture
• Handheld monocular Realme RMX3867
• 1280×720, H.264
Embodied behavior
• Repeated handling across shoes, cloth, cleaner + _shoe, rotate_shoe, apply_cleaner, change_pair, carry_shoes
• Challenges: object identity, state tracking, tool changes + long occlusions
Video sourced from Numo. Numo collects real-world data that Poseidon refines into training-ready datasets.
Long-horizon tasks require more than recognizing an action frame by frame. Models need to track which objects they’ve interacted with, how their state has changed, and what still needs to happen.
12
1
41
4,104
How much speech data is enough to build reliable ASR models?
With Numo Indic Speech, we're adding 3,845+ hours of transcribed audio across Hindi, Telugu, Tamil, and Bengali.
That's roughly 6.3x the labeled hours in Kathbath across the same four languages, giving researchers ing and evaluation.
With Numo Indic Speech, we're adding 3,845+ hours of transcribed audio across Hindi, Telugu, Tamil, and Bengali.
That's roughly 6.3x the labeled hours in Kathbath across the same four languages, giving researchers ing and evaluation.
11
0
41
3,641
Explore the datasets: huggingface.co/datasets/psdn-…
With 353K+ recordings, complete reference transcripts, and granular metadata, these datasets give researchers greater depth to investigate model performance within each language.
With 353K+ recordings, complete reference transcripts, and granular metadata, these datasets give researchers greater depth to investigate model performance within each language.
0
0
4
427
We’re open-sourcing SONAR.
SONAR is Poseidon’s evaluation toolkit for multilingual speech recognition (ASR), built to measure more than whether a transcript matches word-for-word.
It combines WER, CER, semantic similarity, and the Poseidon Score with tools for identifying raphic, and other failure modes, including across low-resource languages.
Get started ↓
SONAR is Poseidon’s evaluation toolkit for multilingual speech recognition (ASR), built to measure more than whether a transcript matches word-for-word.
It combines WER, CER, semantic similarity, and the Poseidon Score with tools for identifying raphic, and other failure modes, including across low-resource languages.
Get started ↓
12
2
56
4,744
Now anyone can use SONAR with their own models and audio, extend it to new languages and datasets, or contribute directly.
SONAR Quickstart Guide:
psdn.ai/blog/open-sour…
SONAR on GitHub (MIT licensed):
github.com/PSDN-AI/SONAR-…
SONAR Quickstart Guide:
psdn.ai/blog/open-sour…
SONAR on GitHub (MIT licensed):
github.com/PSDN-AI/SONAR-…
1
0
5
744
Data in the Wild, ep. 003.
Task context
•- Organize food + containers by type and usage
•- 10 minutes
•- A home in Korea
Capture
•- Handheld monocular Samsung SM-A536N
•- 1920×1080, H.264
Embodied behavior
•- Bottles, jars, packages, bins + other household objects
•- spect_package, place_container, remove_bin, wipe_bottle, rinse_cap, align_items, close_door
•- Challenges: varied object shapes + sizes, spatial organization, handling errors
Video sourced from Numo. Numo collects real-world data that Poseidon refines into training-ready datasets.
Organization tasks require models to reason about more than where an object fits. They involve understanding what an object is, how it’s used, where it belongs, and recovering when things don’t go as planned.
Task context
•- Organize food + containers by type and usage
•- 10 minutes
•- A home in Korea
Capture
•- Handheld monocular Samsung SM-A536N
•- 1920×1080, H.264
Embodied behavior
•- Bottles, jars, packages, bins + other household objects
•- spect_package, place_container, remove_bin, wipe_bottle, rinse_cap, align_items, close_door
•- Challenges: varied object shapes + sizes, spatial organization, handling errors
Video sourced from Numo. Numo collects real-world data that Poseidon refines into training-ready datasets.
Organization tasks require models to reason about more than where an object fits. They involve understanding what an object is, how it’s used, where it belongs, and recovering when things don’t go as planned.
16
0
57
4,337
Our biggest speech data release yet:
Today, we’re introducing our Indic Speech Release on @huggingface: 3,845+ hours of transcribed speech across Hindi, Telugu, Tamil, and Bengali.
• 353K+ recordings
• Longer, real-world speech samples
• Speaker + demographic metadata
• signals
Built to help teams train, evaluate, and find the failures that short-clip benchmarks tend to hide.
Today, we’re introducing our Indic Speech Release on @huggingface: 3,845+ hours of transcribed speech across Hindi, Telugu, Tamil, and Bengali.
• 353K+ recordings
• Longer, real-world speech samples
• Speaker + demographic metadata
• signals
Built to help teams train, evaluate, and find the failures that short-clip benchmarks tend to hide.
47
12
107
9,969
All samples sourced via Numo, Poseidon’s data collection app.
Read the full research + explore the datasets:
• Blog: psdn.ai/blog/building-…
• Datasets: huggingface.co/datasets/psdn-…
Read the full research + explore the datasets:
• Blog: psdn.ai/blog/building-…
• Datasets: huggingface.co/datasets/psdn-…
5
0
12
2,204
Data in the Wild, ep. 002.
Task context
•- Clean a small aquarium + replace water
•- 10 minutes
•- A home in the Philippines
Capture
•- Handheld monocular Samsung SM-A135F
•- 1280×720, H.264
Embodied behavior
•- Two-handed handling across glass, water, tools + small s: remove_decor, transfer_water, rinse_vessel, replace_decor, fit_lid, wipe_surface, return_container
•- Challenges: transparent water, reflective glass, wet-tool handling + long-horizon sequencing
Video collected from Numo, our data collection app that sources real-world data. Poseidon then refines these captures for model training, with a focus on long-tail tasks.
Task context
•- Clean a small aquarium + replace water
•- 10 minutes
•- A home in the Philippines
Capture
•- Handheld monocular Samsung SM-A135F
•- 1280×720, H.264
Embodied behavior
•- Two-handed handling across glass, water, tools + small s: remove_decor, transfer_water, rinse_vessel, replace_decor, fit_lid, wipe_surface, return_container
•- Challenges: transparent water, reflective glass, wet-tool handling + long-horizon sequencing
Video collected from Numo, our data collection app that sources real-world data. Poseidon then refines these captures for model training, with a focus on long-tail tasks.
24
5
53
8,108
Data in the Wild, ep. 001.
The data needed to train robotics is rare, unscripted, and full of long-tail behaviors. Here’s what that looks like in practice.
Task context:
•- Brush the pet, wipe its paws & body, collect loose fur
•- 11 minutes
•- A home in the Philippines
Capular iPhone
•- 1920×1080 H.264
Embodied behavior:
•- One hand stabilizes or repositions the dog while the other brushes
•- Actions: stabilize_pet, brush_coat, reposition_pet, set_down_brush
•- Challenge: Moving occlusion, careful hand manipulation, unpredictable animal motion
The data needed to train robotics is rare, unscripted, and full of long-tail behaviors. Here’s what that looks like in practice.
Task context:
•- Brush the pet, wipe its paws & body, collect loose fur
•- 11 minutes
•- A home in the Philippines
Capular iPhone
•- 1920×1080 H.264
Embodied behavior:
•- One hand stabilizes or repositions the dog while the other brushes
•- Actions: stabilize_pet, brush_coat, reposition_pet, set_down_brush
•- Challenge: Moving occlusion, careful hand manipulation, unpredictable animal motion
10
5
61
6,919
Video sourced from Numo, Poseidon’s data collection app.
Numo sources real-world data that existing datasets don’t capture. Poseidon then processes that data for model training.
Pet brushing + paw wiping: ~19 hours in Numo vs. 0 across 12 public egocentric video datasets we analyzed.
Numo sources real-world data that existing datasets don’t capture. Poseidon then processes that data for model training.
Pet brushing + paw wiping: ~19 hours in Numo vs. 0 across 12 public egocentric video datasets we analyzed.
6
0
9
2,667
Numo on web is winding down as we move everything to mobile. Download and keep contributing.
iOS → apps.apple.com/app/id67835425…
iOS → apps.apple.com/app/id67835425…
2
0
4
1,550
Conversational voice data exposes a subtle failure mode in many ASR pipelines. The metric says the data is bad, but human reviewers say the audio is clear.
Often the issue isn’t transcription quality itself, but that the evaluation stack wasn’t built for turn-taking, silence, ture.
Learn how to solve this in our blog:
Often the issue isn’t transcription quality itself, but that the evaluation stack wasn’t built for turn-taking, silence, ture.
Learn how to solve this in our blog:
21
6
60
7,931