A Review of Communicating Robot Learning During HRI
Authors: Soheil Habibian, Antonio Alvarez Valdivia, Laura H. Blumenschein, Dylan P. Losey Published: International Journal of Robotics Research, 2024
Overview
Surveys the intersection of robot learning algorithms and the interfaces that communicate robot learning state to human teachers. Argues that learning and communication must be designed together — not as separate modules — to achieve genuine human-robot co-adaptation.
The Gap This Addresses
Most LfD research treats the human as a passive data source. This survey asks: what happens if robots communicate what they’ve learned back to the human? Does it change how the human teaches?
Answer: yes — significantly.
Communication Interface Taxonomy
Visual interfaces
- Visualizations of learned reward functions, uncertainty, intended trajectories
- Most common; lowest bandwidth for conveying internal state
Haptic feedback
- Force-reflecting teleoperation (human feels robot resistance or compliance)
- High bandwidth for physical tasks; most informative for manipulation
- Underexplored relative to visual
Auditory channels
- Alerts, sonification of robot confidence or state
- Useful for attention management; rarely used alone
Key Findings
- When robots communicate their learning (not just execute), humans adapt their teaching strategy
- Closed-loop co-adaptation leads to: better human demonstrations, increased human trust, faster skill acquisition
- Haptic feedback during kinesthetic teaching is uniquely valuable because it is in-band with the demonstration interface itself — no attention split required
Case Study
Kinesthetic teaching of a robot arm with multimodal feedback: participants changed the quality and character of their demonstrations based on what the robot communicated, resulting in measurable improvement in learned policy performance.