Demonstration Modality Impact on Imitation Learning (ICRA 2025)

Summary

ICRA 2025 paper by Li, Cui, and Sadigh (Stanford) comparing three demonstration collection methods for robot imitation learning: kinesthetic teaching (physically guiding the robot), VR teleoperation, and spacemouse. Key finding: kinesthetic teaching produces cleaner data but causes fatigue; a hybrid scheme combining a small kinesthetic seed dataset with VR data achieves ~20% better performance than either alone.

Stanford Li、Cui 和 Sadigh 的 ICRA 2025 論文，比較三種機器人模仿學習示範收集方法：動覺教學（物理引導機器人）、VR 遠程操控和空間滑鼠。主要發現：動覺教學產生更乾淨的資料但造成疲勞；結合少量動覺種子資料集與 VR 資料的混合方案，比單獨使用任一模態實現約 20% 更高的性能。

Key Points

• Kinesthetic teaching: highest data quality (consistent, low-noise trajectories) but low scalability due to physical fatigue
• VR teleoperation: higher state diversity, scalable, but noisier action trajectories
• Spacemouse: least intuitive, lowest quality
• Hybrid scheme (small kinesthetic + large VR): +20% performance vs. either modality alone
• Quality dimensions that matter: action consistency (kinesthetic advantage) vs. state diversity (VR advantage)

Insights

• The hybrid finding is a practical recipe for real-world data collection: “seed” quality + “scale” quantity. This directly explains GR-Dexter’s data collection design — they use foot pedals + Manus gloves for scale but the architecture implicitly benefits from the quality that physical guidance provides
• The action consistency vs. state diversity framing is a useful lens for evaluating any demonstration interface: what is the tradeoff between how well it captures expert intent (consistency) and how many situations it covers (diversity)?
• Physical fatigue as the bottleneck for kinesthetic teaching is non-obvious: most papers optimize for data quality, but the human cost of collection is often what limits real-world robot learning programs
• This paper is about 3-DOF tabletop manipulation tasks — the conclusions may not hold for 21-DOF dexterous hands where kinesthetic teaching is physically impossible (too many joints to simultaneously guide)

Connections

• GR-Dexter: VLA for Bimanual Dexterous Robot Control
• RoboCopilot Human-in-the-Loop Imitation Learning
• State of VLA Research at ICLR 2026
• Learning from Demonstration
• Imitation Learning
• Teleoperation

Raw Excerpt

A simple data collection scheme that relies on a small number of kinesthetic demonstrations mixed with data collected through teleoperation achieves the best overall learning performance while maintaining low data-collection effort.