RoboPapers Ep#16 TWIST: Teleoperated Whole-Body Imitation System
Jun 25, 2025
Yanjie Ze, a first-year PhD student at Stanford, dives into the innovative TWIST system, enhancing humanoid robot capabilities through teleoperation. The discussion reveals how human data dramatically improves robot dexterity and addresses challenges in lower body tracking. Yanjie explains the significance of large-scale motion datasets for refining robot movements and explores the complexities of control frameworks. The conversation also highlights advancements in teleoperated robotics, focusing on latency improvements and the potential of full-body engagement for enhanced robotic performance.
AI Snips
Chapters
Transcript
Episode notes
Teacher-Student Learning Framework
- Employ a teacher-student policy framework to distill future motion information into real-time control.
- Use sparse future sampling to balance latency and prediction horizon.
Reward Learning Beats Cloning
- Reward-based learning outperforms behavioral cloning for generalizable whole-body motion control.
- Combining reward optimization with behavioral cloning from mocap yields best results.
Optimize for Low Latency
- Focus engineering efforts on reducing latency in the retargeting stage.
- Achieving under 0.1s latency makes teleoperation feel instantaneous to humans.