Reproduce “TinyAstral: Transferring Self-Evolving Harnesses to Tiny Generalist Manipulation Agents” 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
Kehong Gong, Xiaoxu Zheng, Guanli Hou, Seunghun Lee, Mengqi Guo, Runjia Tan, Zhengyu Wen, Qi Wang and 3 more
TinyAstral improves a robot-planning harness around a frozen model, then trains a 4B student on its reasoning, tool use, and recovery. RoboTwin-50 strict success rises from 33.4% to 63.6% for the teacher and from 36.2% to 62.8% for separately trained students across harness versions. In one setup, the student uses 4.9× less accelerator time per episode, not a guarantee elsewhere.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.