Event
RoboCoach: World-Model Localization of Weak Subtasks for Targeted Teaching
1 reports · 1 sources
The team of Zhang Changshui and Cui Sen at Tsinghua University (the research is reported to have been completed by teams from Renmin University of China and Tsinghua University) proposes an embodied intelligence system, RoboCoach: it first uses a world model, CoachWorld, to run long-horizon tasks in a closed loop in imagination, locates the first weak subtask that times out, and then requests human teaching specifically for it, updating only the corresponding skill expert. Under a fixed demonstration budget, with each method starting from the same initial policy and undergoing three rounds of updates, the real-robot success rate rose from 13.3% to 75.0% on Franka and from 40.0% to 83.8% on AgileX; the uniform-collection baselines were 30.0% and 47.5%, respectively. The average success rate on four unseen new task routes was 35.0%, compared with 0% for the shared-policy baseline. Each round adds 100 qualified subtask demonstrations in simulation and 50 per real-robot platform; real-robot testing covers 3 tasks on the Franka Research 3 and 4 tasks on the AgileX dual-arm platform.
Report timeline
- RoboCoach: Using World Models to Locate Weak Subtasks and Improve Long-Horizon Task Success Through Targeted Teaching
Zhang Changshui and Cui Sen's team at Tsinghua University propose RoboCoach, an embodied intelligence system that uses the world model CoachWorld to run closed-loop trials of long-horizon tasks in imagination, identify the first weak subtask that times out, and request targeted human teaching for it, updating only the corresponding skill expert. Under a fixed demonstration budget, the success rate rises from 13.3% to 75.0% on Franka and from 40.0% to 83.8% on AgileX, while the uniform collection baseline reaches only 30.0% and 47.5%, respectively. On four unseen new task routes, the average success rate is 35.0%, compared with 0% for the shared-policy baseline.
机器之心 · 微信公众号 ·