RoboPapers Ep#10 Humanoid Policy ~ Human Policy
Jun 10, 2025
In this discussion, Roger Qiu, a first-year PhD student at UCSD focused on humanoid policy in robotics, delves into some mind-bending topics. He shares insights on the challenges of teleoperating humanoid robots and how utilizing human data can significantly enhance their skills. The conversation also highlights the potential of mixed reality in robot learning, the importance of task-specific data, and the quest for optimizing data collection in robotic tasks. With a touch of humor, Roger explores the complexities of robotics in the ever-evolving Roboverse.
AI Snips
Chapters
Transcript
Episode notes
Efficient Human Data Collection
- Collecting human demonstrations requires no specialized skills and minimal training.
- Around 30 to 35 man-hours produced about 25 hours of high-quality human demonstration data.
Scalable Human Data Collection
- Human demonstrations enable fast, diverse, and scalable data collection without robots or computers.
- Using lightweight VR apps like on Apple Vision Pro allows data capture even outdoors or in natural settings.
Unified Human-Robot Representation
- A unified human-centric representation handles both human and humanoid robot motions.
- The policy operates without explicit inverse kinematics; retargeting is managed in data preprocessing.