LIBERO-CF64k / README.md
OpenWAM-Stanford's picture
Configure CC BY 4.0 data license and retain code/upstream notices
36d736f verified
|
Raw History Blame Contribute Delete
1.78 kB
metadata
license: cc-by-4.0
language:
  - en
size_categories:
  - 10K<n<100K
tags:
  - robotics
  - libero
  - counterfactual-learning
  - inverse-dynamics
  - forward-dynamics
  - video
  - openwam

OpenWAM LIBERO-CF64k

Project | Code

Learning what happens beyond the demonstration. LIBERO-CF64k contains 64,000 counterfactual motion segments across the ten LIBERO-Long (LIBERO-10) tasks. Different action sequences are executed from shared starting states, providing examples of how robot motion changes the scene beyond successful expert demonstrations.

Interventions include gripper opening and closing, directional motion, rotation, pauses, changes in speed, and randomized controls. The collection supports learning both inverse dynamics, which maps visual motion to actions, and forward dynamics, which predicts the effects of actions.

What's Included

  • 64,000 segments, balanced at 6,400 per task, from 10,000 branching contexts.
  • Agent-view and wrist-camera RGB video, recorded actions, and robot proprioception.
  • Boundary simulator states and task descriptions.
  • A uniform 20 Hz format: 128 controls and 129 observations per 6.4-second segment.

The release provides raw data, a reader, and integrity checks. See the project page and code repository for the method and tools.

Acknowledgments

The collection builds on the tasks and simulation environments from LIBERO.

License

The dataset is released under CC BY 4.0. The reader code uses AGPL-3.0-only. Attribution and upstream terms are recorded in NOTICE.