Reproduce “ALIGN: Aligning Condition Estimation and Expression in Advantage-Conditioned Behavior Cloning” 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
Zechen Wang, Zhenqi Qiu, Houlong Xiong, Suohang Zhang, Yiyu Ren, Wanting Xu, Hongfei Niu, Chengyang He and 3 more
ALIGN trains a policy to use retry-containing demonstrations by assigning binary progress labels and injecting them throughout the action-generating network. It compares same-trajectory observations to assign labels directly. On ARX X5 three-bowl stacking, it succeeds in 90.0% of 30 rollouts, versus 56.7% for the prior AC-BC baseline; these point estimates do not establish independent component effects.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.