The Role of Conventions in Adaptive Human-AI Interaction
summary
Argues that encoding partner behavior into a learned low-dimensional latent intent Z via an autoencoder-style prediction loss and feeding it into a reinforcement learning loop yields faster, more interactive coordination; implements LiLI on real robots for air hockey and reports about 90% success after four hours of training (drops to ~70% with humans). Extends to a stabilizing reward (SILY) that reduces nonstationarity and to a decentralized UCB + follower model that achieves logarithmic regret.
AI-written summary · about
chapters (20)
- 0:00Introduction
- 3:05Autonomous Driving
- 5:20Changing Lanes
- 8:50Coordination Tasks
- 13:35Understanding the Action Space
- 32:45Nonstationary Agents
- 33:35Nonstationarity
- 35:00Reducing Nonstationarity
- 35:25Representation Learning
- 36:25Merrill
- 37:30Stable
- 37:45Lilly
- 38:55Summary
- 41:35Modular Architecture
- 42:45Collaborative Multiarm Bandits
- 46:05Robotics Example
- 48:40Thank you
- 49:10Humans are smarter than robots
- 51:15Collaborative tasks
- 53:15Intrinsic motivation
related talks
26mAI Agents and the Future of Human-AI Collaborative Intelligence
08mHuman-AI Collaboration in Cooperative Games: A Study of Playing Codenames with an LLM Assistant
43mAI Has Entered The Chat: Humans, AI, and the Art of Collaboration
36mOn the Road to Artificial General Intelligence
42mRethinking AI Alignment and Evaluation under Heterogeneous Preferences
16mReal-World Robotics as the Next Frontier for AI?
35mIndustrial Applications of Reinforcement Learning
32mThe Road to General AI