Reproduce “Cross-Platform Control for Autonomous Surface Vehicles via Adaptive Reinforcement Learning” 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
Ruiheng Jiang, Thomas Bi, Raffaello D'Andrea, Aswin Ramachandran
An adaptive reinforcement learning framework from ETH Zurich enables autonomous surface vehicles (ASVs) to perform zero-shot cross-platform trajectory tracking. This single policy, trained on a lightweight 3-DoF model, learns to adapt to diverse vessel dynamics online, achieving comparable performance to platform-specific tuned controllers on physical ASVs and reducing position errors by up to 58% over non-adaptive generalist policies.
Key insights
- Formulating cross-platform control as a partially observable Markov decision process allows the policy to infer unknown dynamics from real-time interaction history.
- Training a teacher policy with privileged dynamics information and then distilling this knowledge into a student adapter provides an effective mechanism for online adaptation.
- A lightweight 3-DoF model with broad dynamics randomization can effectively train a generalist policy, circumventing the need for complex, high-fidelity simulators and completing training in approximately 30 minutes on a single consumer GPU.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.