Reproduce “Value as Memory: Online Value-Guided Post-training for General Robot 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
Wenye Yu, Yijin Chen, Kaiyi Xiong, Wenbo Han, Wendi Chen, Yang Jin, Han Xue, Jun Lv and 2 more
VaM gives a robot’s action policy recent value predictions as memory during long-horizon manipulation. The policy combines this cache with visual and proprioceptive history, while online training updates both models from deployment experience. Across three real-robot tasks, caching improved success by a mean 29.4 percentage points over no caching, under methods’ respective training setups.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.