podolinsky2025
GR00T-N1.7 LIBERO-X backbone features (90-task fine-tune, LEVEL1-3)
Aligned rollouts of a fine-tuned GR00T-N1.7 model on the LIBERO-X simulator across 90 tasks at three difficulty levels, with extracted layer-16 backbone features, video, and action data. Contains 1,800 episodes (600 per level) with vision-language representations, proprioceptive state features, and executed actions.
Downloads16
Episodes1800
Why This Matters for Physical AI
Provides fine-tuned vision-language-action backbone representations and policy rollouts on a distribution-shifted benchmark to enable research on robustness, transfer learning, and interpretability of foundation models for robotic manipulation.
Technical Profile
- Modalities
- videotabularlanguageproprioception
- Robot Embodiments
- humanoid
- Action Space
- end_effector_delta
- Environment
- simulation
- Task Types
- manipulationpick_and_place
- Episodes
- 1800
- Data Format
- npz
- Annotation Types
- language_instructionsreward_labels
Access
Need custom video data?
Claru builds purpose-built datasets for simulation applications with dense human annotations and quality assurance.
Request a Sample Pack