Reproduce “A Controlled Study of Models, Skills, Feedback, and Curriculum in Agentic Robotics” 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

Kaiyuan Chen, Xinyue Zhu, Shuangyu Xie, Uksang Yoo, Thi Thu Phuong Luong, Marcus Gualtieri, Christian Juette, Liu Ren and 2 more

The authors study how feedback and spatial representations affect agents’ revisions of executable robot policies in simulation. They compare episode-, motion-, and skill-level reports, alongside coordinate and part annotations. Per-object held-out success averages 52% with per-skill feedback versus 25% per-episode feedback, but Syringe Packing favors per-episode feedback, and results do not establish real-world transfer.

Summary by alphaXiv.

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