Loop data for robot policy improvement.
Map high-signal robot data, clean demonstrations, evaluate checkpoints, and turn failed rollouts into the next training set.
Loop data
find → structure → test
From useful data to measurable failures.
Free entry
Curated robot + ego dataset map.
Physical AI needs data that captures the loop.
What happened, what failed, and what data comes next.
Find the right data
Access, license, modality, embodiment, signals.
Trace the episode
Observation, action, feedback, correction.
Find the break
Subtask, environment, feedback point.
Replay the gap
Hard cases back into training.
Find loop-worthy data.
Robot, ego, teleop, and human demonstration sources compared in one catalog.
- Access / license / format
- Ego4D, LeRobot, RLDS, HDF5
- Video, action, pose, IMU, hands
Manipulation, navigation, humanoid tasks, and failure datasets.
Start with data. Build the measured loop.
Catalog, Data, and Bench are the active path. Synth remains the planned failure-replay step.
- 01
OpenBot Catalog
Find loop-worthy data
Access, format, embodiment, signal.
Explore Catalog - 02
OpenBot Data
Structure loop traces
Deduplicate, detect drift, preserve feedback.
Explore Data - 03
OpenBot Bench
Locate failure loops
Track success, recovery, and worst cases.
Explore Bench - 04
OpenBot Synth
Planned · failure replay
A future service for replaying measured failures in simulation.
View roadmap - 05
OpenBot API
Automation surface
Async workflows for CI, runners, and tools.
Explore API