本文由 AI 分析生成
建立時間： 2026-05-30 來源： https://arxiv.org/abs/2503.07771
Summary
RoboCopilot addresses the covariate shift problem in imitation learning by enabling seamless real-time control switching between a human teacher and an autonomous robot policy on a bimanual manipulator. Rather than collect demonstrations passively and train offline, the human intervenes when the policy fails, with each intervention becoming new on-policy training data.
RoboCopilot 解決模仿學習中的 covariate shift 問題：人類老師和自主策略共享雙臂機器人的控制權，在策略失敗時無縫介入並提供示範。這實際上是硬體上的 DAgger 演算法，人類提供 on-policy 修正而非被動的離線示範收集。合規力回饋介面讓老師能感受機器人感受到的力量。
Prerequisites
- DAgger — Dataset Aggregation: the theoretical framework this implements in hardware
- Covariate shift — why offline-trained policies fail at deployment: training states != states the policy visits
- Bilateral teleoperation — force-reflecting interface enabling natural human intervention
Core Idea
Traditional LfD collects demos passively then trains offline — the policy never sees the states its own imperfect execution creates. RoboCopilot closes this loop: deploy initial policy, human intervenes when policy drifts, intervention provides training data from states the policy actually visits. The bilateral (force-reflecting) interface is critical — the human feels robot forces enabling precise corrections.
Results
| Metric | Value |
|---|---|
| Setup | Bimanual manipulation, sim + hardware |
| Learning speed | Faster skill acquisition vs. passive collection |
| Control switching | Real-time, mid-task, no interruption required |
Limitations
- Author-stated: requires continuous human monitoring during execution
- Unstated: bilateral teleoperation hardware is expensive; force-reflecting rigs add complexity in unstructured environments
Reproducibility
- Code: Not mentioned in clipping
- Dataset: Not released per clipping
- Compute: Standard manipulation policy training
Insights
The seamless handoff is the key engineering contribution — prior work required task restarts for interventions. The force-reflecting interface generates richer training signal than vision-only teleoperation by capturing the contact dynamics the policy needs to learn. This architecture is most valuable for contact-rich tasks where covariate shift is severe.
Connections
Raw Excerpt
“RoboCopilot approach: Deploy initial policy → Human intervenes when policy fails → Intervention becomes new training data → Policy improves with each correction. This is DAgger in hardware.”