Kavin606062026apache-2.0
ISR vs uniform downsampling vs full data — closed-loop success on ALOHA human teleop demos
A closed-loop evaluation comparing Information-Standardized Trajectory Resampling (ISR) against uniform downsampling and full data on ALOHA sim transfer cube tasks using ACT policy. Tests whether ISR improves sample efficiency and validates findings on public human teleop demonstrations.
Downloads13
Episodes50
Why This Matters for Physical AI
Demonstrates that trajectory resampling based on information content rather than temporal uniformity improves sample efficiency in imitation learning, enabling policies trained on 50% of frames to match full-data performance—a critical finding for efficient robot learning from demonstrations.
Technical Profile
- Modalities
- rgbproprioception
- Robot Embodiments
- ALOHA
- Action Space
- joint_positions
- Environment
- simulation
- Task Types
- manipulationpick_and_place
- Episodes
- 50
- Data Format
- LeRobot
- Annotation Types
- action_labels
- License
- apache-2.0
Access
Need custom rgb data?
Claru builds purpose-built datasets for simulation applications with dense human annotations and quality assurance.
Request a Sample Pack