RoboCoach: World Models as Active Coaches for Compositional Robot Skills
迈向Physical RSI第一步，具身智能系统RoboCoach如何高效主动纠错？
RoboCoach is a system that uses world models to enable efficient active correction in embodied reinforcement learning. By decomposing long tasks into subtasks and directing human demonstrations to specific experts, RoboCoach achieves significant performance improvements on real robots like Franka and AgileX, with success rates increasing from 13.3% to 75.0% and 40.0% to 83.8%, respectively. The system's RIDI framework (Route—Imagine—Diagnose—Improve) enables targeted skill refinement, reducing the need for extensive human intervention.
Source: Machine Heart — Robotics on WeChat · Read original article ↗
English translation unavailable. Choose the original article to read it.
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
Reported numbers
Reported success rate
13.3%
View original evidence
RoboCoach 在有限人工示范预算下，将 Franka 的成功率从 13.3% 提升至 75.0%，AgileX 从 40.0% 提升至 83.8%，相比起基线方法有显著效率提升。
Open source S8Reported success rate
75%
View original evidence
RoboCoach 在有限人工示范预算下，将 Franka 的成功率从 13.3% 提升至 75.0%，AgileX 从 40.0% 提升至 83.8%，相比起基线方法有显著效率提升。
Open source S8
- paper: not_reported
What remains unknown
Not established in the collected evidence: Environment, Control, Data origin.
Reported performance applies to the described task. It does not establish general autonomy or deployment readiness.
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
RoboCoach 在有限人工示范预算下，将 Franka 的成功率从 13.3% 提升至 75.0%，AgileX 从 40.0% 提升至 83.8%，相比起基线方法有显著效率提升。
Open source S8
RoboCoach 通过复用子任务的技能专家，在四个已有技能重新排序、跨任务拼接或直接接长的 held-out 任务上， 取得 35.0% 的平均成功率 ，而对应的共享策略基线为 0% 。
Open source S15
Source：Machine Heart — Robotics on WeChat · mp.weixin.qq.com