Releases·Tool Release
Physical AI's Next Bottleneck Is Finding the Right Video — Rafael Levi, Bright Data
In a conference talk, Bright Data's Rafael Levi argues physical AI's bottleneck is action-labeled video, not models. LLMs train on trillions of words while robotics has only about a million robot-action videos, and staged recordings introduce bias. Bright Data's answer is a "search first, collect second" API that indexes over a billion web videos by actions and returns timestamped, scored clips.
28 engagement1 sourceThu, Sep 24, 2026, 02:00 PM
AI Overview
- Rafael Levi of Bright Data says robotics has only about a million robot-action videos, while the web holds billions of hours of real people handling objects
- NVIDIA discards about 96% of downloaded video for Cosmos training and Stable Video Diffusion discards 74%, wasting compute, bandwidth and storage
- Meta trained on roughly a million hours of public video and needed only 62 hours of real robot data to control a robot
- Bright Data indexes over a billion videos by actions, not titles, returning trimmed clips with timestamps, match scores and frame counts via API
- Paying people to record staged actions produces biased data because people don't perform actions naturally when told to
Related
Releases·Tool Releasefeatured
Figure.AI releases Index, a 16M-video robot dataset with paid data collection
Figure.AI released Index, a robot dataset containing 16 million videos, described as the largest and most diverse of its kind. The company is now paying people to record their daily tasks to expand the dataset, which is intended to improve robot learning.
Figure.AI released Index, a robot dataset containing 16 million videos, described as the largest and most diverse of its kind. The company is now paying people to record their daily tasks to expand the dataset, which is intended to improve robot learning.
410 engagement1 sourceredditTue, Aug 25, 2026, 07:20 PM
Social·Community
Figure.AI INDEX, the video dataset for humanoid robots contributed by people, is growing at a rate of 2 million per week
Figure.AI's INDEX humanoid robot video dataset grows by 2 million clips per week
Figure.AI's INDEX humanoid robot video dataset grows by 2 million clips per week
48 engagement1 sourceredditFri, Sep 4, 2026, 07:16 PM
Developments·Paper
Pegasus framework bridges embodiment gap by translating human videos into robot training data
Pegasus is a low-resource framework that converts human demonstration videos into robot-learnable data using a graph-based intermediate representation. It extracts a Task Graph from human videos, then transforms it through Affordance and Constraint mapping to generate physically grounded robot experiences. This addresses the key bottleneck of data scarcity for embodied AI without requiring expensive robot-specific data collection.
Pegasus is a low-resource framework that converts human demonstration videos into robot-learnable data using a graph-based intermediate representation. It extracts a Task Graph from human videos, then transforms it through Affordance and Constraint mapping to generate physically grounded robot experiences. This addresses the key bottleneck of data scarcity for embodied AI without requiring expensive robot-specific data collection.
0 engagement1 sourcearxivWed, Jul 29, 2026, 01:35 PM
Social·Community
Redditor argues selling real jobsite training footage to robotics firms is the next opportunity in physical AI
A Reddit post argues that physical AI needs real footage of people performing trade and jobsite work, unlike generative AI trained on text, code, and video. The author contends that whoever sells or captures this training data for robotics companies will lead the industry, pointing to Mecka AI as an existing example.
A Reddit post argues that physical AI needs real footage of people performing trade and jobsite work, unlike generative AI trained on text, code, and video. The author contends that whoever sells or captures this training data for robotics companies will lead the industry, pointing to Mecka AI as an existing example.
10 engagement1 sourceredditWed, Sep 2, 2026, 08:31 PM
Developments·Paper
KnowDemo: Knowledge-Guided Robot Demonstration Generation from Human Videos
A new arXiv paper proposes KnowDemo, a framework that extracts structured manipulation knowledge from human videos using a vision-language model to generate diverse, verified robot demonstrations in simulation. It addresses limits of motion-reference adaptation by allowing alternative contact strategies and subtask orders, and the authors fine-tune the pretrained π0.5 model on the generated data, achieving sim-to-real transfer across three tasks.
A new arXiv paper proposes KnowDemo, a framework that extracts structured manipulation knowledge from human videos using a vision-language model to generate diverse, verified robot demonstrations in simulation. It addresses limits of motion-reference adaptation by allowing alternative contact strategies and subtask orders, and the authors fine-tune the pretrained π0.5 model on the generated data, achieving sim-to-real transfer across three tasks.
0 engagement1 sourcearxivFri, Sep 18, 2026, 02:19 AM