Automatic Curriculum Design for Zero-Shot Human AI Coordination
Date: 26th June 2025
Key Points
- Curriculum is set by ranking learnability according to normalised group return
- Similar to PLR (link when home), randomly either creates new level or replays old one
- Human AI coordination is achieved by having a very diverse set of agents in the population such that a human will be similar to some.
- Forces agent to learn with worst team mate as this is the hardest.