Reproduce “AnyMorph: Learning Transferable Polices By Inferring Agent Morphology” 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
Brandon Trabucco, Mariano Phielipp, Glen Berseth
AnyMorph trains one variable-size reinforcement-learning policy for agents with different sensor and actuator dimensions. It represents each dimension as a token, embeds observations and actions, and uses a sequence-to-sequence Transformer; token IDs remain manually assigned. On held-out simulated morphologies, it improved zero-shot average return 16% overall and 32% on a 32-morphology task, but not across unseen families or real robots.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.