Robots and physical AI ultimately learn from data. Teleoperation logs collected by people driving real robots, synthetic data churned out by physics engines in simulators, tactile data capturing the pressure and slip at a fingertip, and behavior data accumulated at national scale. A robot foundation model's competence is decided by the scale and quality of this data. This hub gathers the writing on the data robots learn from itself — how datasets are structured and released, their quality and standards, and provenance, the piece still missing.
Pebblous reads this data through one lens: not making more data, but making proof-attached data. Most public datasets keep the frames, states, and actions but discard the proof — contact force, generation history, physical consistency. Filling that gap with data-quality diagnosis, refinement, and certification is the work of DataClinic.
A landscape comparing DROID, Open X-Embodiment, GR00T, RoboCasa, MimicGen, and LIBERO by scale, license, and format. Collection splits six ways, but distribution converges on a single LeRobot format — and all six leave provenance unrecorded. (Overview of this hub.)
Kyung Hee University's PRISM generates digital demonstration data from a single photo of a workspace plus a natural-language instruction, sidestepping the data bottleneck in robot policy learning.
NVIDIA NuRec reconstructs real footage into simulator assets via 3D Gaussian Splatting — a real-to-sim pipeline that moves reality into the simulator.
Alibaba DAMO Academy's RynnWorld-Teleop synthesizes robot-perspective demonstration video from human hand-motion streams alone, without a physical robot.
Ai2's MolmoAct 2 released not just model weights but 720 hours of bimanual-manipulation robot data, code, and evaluation in full — a new bar for open robot data.
AGIBOT WORLD 2026 Theme 2 records not successful demos but the texture of contact — misses, collisions, slips — with tactile sensors. The physics vision alone can't teach.
AgiBotWorld 2026 doesn't discard failed robot demos; it annotates them with error_cause and restorable fields — a rare case of keeping failure history as data.
Humanoid robots and foundation models drew billions, but tactile data capturing pressure and slip drew only millions. The gap in data that capital passed over.
RoboScape enforces physical validity from the moment synthetic data is generated, matching the performance of 200 real videos with 200 synthetic ones.
A dissection of the technique family that audits, after the fact, whether video from Cosmos and Sora is physically correct — how to verify the physical consistency of synthetic data.
Combining 3D Gaussian Splatting with NVIDIA Isaac Sim reshapes the robot synthetic-data pipeline — the frontier of synthetic data generation for VLA training.
How Isaac Sim delivers 1,000x faster training and GR00T Blueprint generates 780K robot trajectories in 11 hours — the mechanics of synthetic data mass production.
With OpenUSD 1.0 standardization, NVIDIA Omniverse became the data-standard layer for physical AI. Even at 780K synthetic trajectories, the quality-assurance gap remains.
Humanoid robots are surging, yet robot experience isn't accumulating. Robot data standardization seen through the ISO/WD 26264-1 draft released in June 2026.
The government named the bottleneck of physical AI not chips or models but "behavior data," and why it plans to build behavior-data training centers across five regions first.
Korea's financial and industry ministries allocated ₩16 trillion in physical-AI policy finance across six industries. Money flows to factories and hardware, but the data gap remains.
However much the data robots learn from grows, if there's no evidence that the data is physically correct, the trust bottleneck stays. See how Pebblous approaches proof-attached data at DataClinic, where data-quality diagnosis, refinement, and certification come together.