Reproduce “MEGA-DAgger: Imitation Learning with Multiple Imperfect Experts” 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
Xiatao Sun, Shuo Yang, Mingyan Zhou, Kunpeng Liu, Rahul Mangharam
MEGA-DAgger, developed by researchers at Yale and UPenn's Embedded Systems Laboratory, introduces an imitation learning algorithm designed to learn from multiple imperfect experts. It produces a policy that surpasses individual expert performance in autonomous racing, demonstrating a 13.6% average improvement in collision avoidance and a 13.2% average improvement in overtaking compared to the collective expert performance.
Key insights
- It is possible to synthesize a policy that not only learns from but also improves upon the performance of multiple individual, imperfect human experts.
- Actively filtering out unsafe expert demonstrations and intelligently resolving conflicting expert advice are critical for developing robust and safe autonomous policies.
- The approach, by balancing safety and performance metrics, can generalize across various autonomous systems beyond racing.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.