How to Train Your Robots? The Impact of Demonstration Modality on Imitation Learning
Authors: Haozhuo Li, Yuchen Cui, Dorsa Sadigh (Stanford) Venue: ICRA 2025 arXiv: 2503.07017
Abstract
This study examines how different methods of providing robot demonstrations affect learning performance and user experience. The authors compare three demonstration modalities: kinesthetic teaching (physically guiding the robot), VR controller teleoperation, and spacemouse teleoperation. Kinesthetic teaching produces data that leads to higher policy performance, but users avoid it for large-scale data collection due to physical demands. The proposed hybrid approach combining small amounts of kinesthetic data with teleoperation data achieves 20% average performance improvements.
Methodology
Robot: 7-DoF Franka Emika Panda with Cartesian Impedance Controller
Three Tasks:
- Open Drawer (constrained linear motion)
- Flip Glass (large rotation near joint limits)
- Push Sanitizer (high contact force required)
Data Collection: 100 trajectories per modality per task, single expert demonstrator.
Modality Details:
- Kinesthetic Teaching: Human physically guides robot arm, poses recorded passively then replayed to recover commanded actions
- VR Teleoperation: Oculus Quest controller tracks hand motion
- SpaceMouse Teleoperation: Button-based velocity control
Policy Training: Diffusion-based Behavioral Cloning, predicts 16-step action sequences, executes first 8.
Data Quality Metrics: Action variance among K-nearest neighbors in proprioceptive state space (action consistency) + state diversity estimation.
User Study: 12 participants (6 experts, 6 non-experts), NASA-TLX surveys.
Results
Policy Success Rates:
| Task | Kinesthetic | VR | SpaceMouse |
|---|---|---|---|
| Open Drawer | 95% | ~70% | ~65% |
| Flip Glass | 70% | ~55% | ~50% |
| Push Sanitizer | 35% | ~55% | ~50% |
Hybrid Collection (30 Kinesthetic + 70 VR):
- Open Drawer: 100% (↑ from 95%)
- Flip Glass: 75% (↑ from 70%)
User Experience:
- Kinesthetic: 最直覺（學習曲線短，p=0.038），但體力負擔最高（p=0.016）
- 儘管評價最好用，大多數參與者仍偏好用遙操作做大規模收集
Data Quality Finding:
- Kinesthetic → 高 Action Consistency（低動作變異數）
- Teleoperation → 高 State Diversity（狀態多樣性更好）
Key Findings
- Kinesthetic > Teleoperation 的資料品質（非接觸任務），但在需要接觸力的任務中表現差，因為 Replay 機制對 Jerkiness 敏感
- 使用者偏好和資料品質解耦：客觀品質最好的方法不是使用者願意長期使用的方法
- 混合方案最佳：少量 Kinesthetic（30%) + 多量 Teleoperation（70%）達到最高成功率，互補各自的弱點
- 任務相依性強：接觸力任務（Push Sanitizer）中 Kinesthetic Replay 出現 Jerkiness，反而不如遙操作
Limitations
- 單一示範者，風格變異有限
- Kinesthetic 缺乏 Force Sensor 整合，需要啟發式補償
- 非專家資料雜訊過大，無法有效分析
- 混合實驗只測了兩個任務
Connections
- Kinesthetic Replay 的 Jerkiness 問題與 Admittance Control 架構密切相關
- 混合收集策略可與 MimicGen 自動擴增結合：先 Kinesthetic 收少量高品質 seed，再擴增
- State Diversity vs Action Consistency 的取捨對應 Behavior Cloning 的 Covariate Shift 問題