Reproduce “Environment Probing Interaction Policies” 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
Wenxuan Zhou, Lerrel Pinto, Abhinav Gupta
The authors introduce EPI, a separate probing policy that briefly interacts with a simulated environment and produces an embedding for a task policy. The embedding is trained to improve next-state prediction rather than maximize task reward. In held-out simulated Hopper environments, EPI achieved 1303 ± 173 accumulated reward per episode versus 1143 ± 156 for History Policy, within two MuJoCo simulation tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.