Summary
IJRR survey by Habibian and Losey arguing that robot learning and human-facing communication interfaces must be designed together, not separately. When robots convey what they’ve learned (via visual, haptic, or auditory feedback) back to human teachers, humans adapt their teaching — creating a closed-loop co-adaptation system that produces better demonstrations, faster skill acquisition, and higher trust.
Habibian 和 Losey 的 IJRR 調查認為，機器人學習和面向人類的通訊介面必須一起設計，而非分開。當機器人通過視覺、觸覺或聽覺回饋向人類教師傳達已學到的內容時，人類會調整他們的教學方式——創造出一個閉迴路共同適應系統，產生更好的示範、更快的技能習得和更高的信任度。
Key Points
- Three communication modalities: visual (visualized reward/uncertainty), haptic (force feedback), auditory (alerts/sonification)
- Most current LfD treats humans as passive data sources — this survey asks what happens when the robot communicates back
- Closed-loop result: humans change their teaching strategy in response to robot feedback → better demonstrations → better policy
- Haptic feedback is uniquely in-band during kinesthetic teaching — no attention split, maximum information density
- Co-adaptation is bidirectional: robot improves from human corrections, human improves from robot feedback
Insights
- The core reframe is profound: LfD is not data collection, it’s a teaching relationship. The design of the teaching interface shapes both what data is collected and how the human conceptualizes the robot’s learning
- Haptic feedback during kinesthetic teaching is the most underexplored modality — the reason is instrumentation cost (force-reflecting arms are expensive), but it’s the most natural channel because it’s already present in the physical act of guidance
- The trust dimension is important: when humans understand what the robot has learned, they calibrate their interventions appropriately — they don’t over-teach (wasting effort) or under-teach (leaving gaps). This is explainability in a physical, interactive form
- This paper is the theoretical complement to RoboCopilot: RoboCopilot builds the hardware for interactive teaching; this survey provides the framework for why the communication component matters as much as the control component
Connections
- RoboCopilot Human-in-the-Loop Imitation Learning
- How to Train Your Robots Demonstration Modality
- Human-Robot Interaction
- Learning from Demonstration
- Haptic Feedback
- Explainable AI
- Trust in HRI
Raw Excerpt
When learning and communication are developed together, the resulting closed-loop system can lead to improved human teaching, increased human trust, and human-robot co-adaptation.