Reproduce “R2BC: Multi-Agent Imitation Learning from Single-Agent Demonstrations” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Connor Mattson, Varun Raveendra, Ellen Novoseller, Nicholas Waytowich, Vernon J. Lawhern, Daniel S. Brown
Round-Robin Behavior Cloning (R2BC) enables multi-agent imitation learning by allowing a single human to provide sequential demonstrations to individual agents, outperforming oracle joint behavior cloning in simulation and achieving up to 5.9x better performance than centralized behavior cloning on physical robots for cooperative tasks.
Key insights
- The online learning paradigm of R2BC, coupled with the active presence of autonomous teammates, naturally introduces observational variability that effectively reduces covariate shift, similar to more complex online imitation learning techniques.
- Decentralization in both policy execution and demonstration collection is crucial for developing practical and robust multi-agent imitation learning systems that align with human demonstration capabilities and real-world deployment constraints.
- Training cooperative multi-agent policies from realistically gathered single-agent human demonstrations can yield performance comparable to or exceeding that achieved with idealized, oracle-generated joint demonstrations.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.