amsks2026MIT
FactoredFB Testbed — Metaworld Play-Resample
Offline play datasets for Metaworld push and pick-place tasks collected with goal resampling every 44 steps to support goal-conditioned reinforcement learning with hindsight relabeling. Contains 1,000,000 transitions per task with 39-dimensional observations and 4-dimensional actions.
Downloads0
Episodes2000 trajectories per task (4000 total)
Why This Matters for Physical AI
This dataset enables research in offline goal-conditioned reinforcement learning with hindsight relabeling by providing densely relabeled play data that matches the structure of OGBench benchmarks, allowing direct comparison of critic architectures and learning algorithms across standardized environments.
Technical Profile
- Modalities
- proprioception
- Robot Embodiments
- Sawyer
- Action Space
- joint_velocities
- Environment
- simulation
- Task Types
- manipulationpushpick_and_place
- Episodes
- 2000 trajectories per task (4000 total)
- Data Format
- npz
- Annotation Types
- reward_labelsgoal_labels
- License
- MIT
Access
Need custom proprioception data?
Claru builds purpose-built datasets for simulation applications with dense human annotations and quality assurance.
Request a Sample Pack