Reproduce “HARP-VLA: Human-Robot Aligned Representation Learning for Vision-Language-Action Model” 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
Xiang Zhu, Puzhen Yuan, Yichen Liu, Jianyu Chen
HARP uses paired human–robot demonstrations to align visual features and learn latent actions from paired and unpaired videos. Its latent-action model predicts motion across embodiments, while a robot-trained action head maps those representations to executable commands. On CALVIN ABC→D, HARP-VLA averages 4.481 completed subtasks, versus 3.917 for OpenVLA-OFT; these are system-level comparisons, not controlled architectural tests.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.