Reproduce “CounterAlign: Counterfactual Supervision for Vision-Language-Action Models” 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

Haru Kondoh, Kei Ota, Asako Kanezaki, Yueh-Hua Wu

CounterAlign reuses successful robot demonstrations by pairing actions with mismatched instructions, learning which instruction-action combinations do not fit. It combines discriminator scores into a reward for offline reinforcement learning, improving a policy without new robot rollouts or reward annotations. On LIBERO-Object position perturbations, it achieves 0.51 success versus 0.19 for π₀.₅; this result does not establish general language understanding.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.