Learning Socially Appropriate Robo-waiter Behaviours through Real-time User Feedback
summary
A custom crowdsourced dataset (8–11 annotators per position) maps appropriateness on a 9x9 grid, used to pretrain RL agents; DQN outperforms A3C and DQN+HER when trained over 20,000 episodes. Real-robot HRI with 21 participants uses four within-subject conditions: random, pretrained, explicit feedback via keyword spotting, and implicit feedback via facial valence. Both adaptive conditions increase enjoyment, sociability and appropriateness; explicit feedback yields the largest gains in sociability and adaptability, implicit best matches users’ felt responses.
AI-written summary · about
chapters (7)
- 0:00Introduction to Robo-waiters
- 1:06Simulating restaurant environment
- 3:16Analyzing user feedback results
- 5:03Reinforcement learning approach
- 6:09Human-robot interaction study
- 8:54Study findings and conclusions
- 10:00Acknowledgments and thanks
related talks
28mPreference Learning from Minimal Human Feedback for Interactive
15mSensing and Handling Engagement Dynamics in Human-Robot Interaction Involving Peripheral ...
14mHuman-in-the-Loop Reinforcement Learning
16mReal-World Robotics as the Next Frontier for AI?
18mComparing Social Robot, Screen and Voice Interfaces for Smart-Home Control
06mInterview with Jong Lee
14mTowards Unobtrusive Physical AI: Augmenting Everyday Objects with Intelligence and Robotic Moveme...
19mRobots as professional Chefs