Reproduce “From VLM Planners to Robotic Agents: Self-Evolving Harnesses for History-Dependent Manipulation” 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
Tingwen Du, Haoyu Guo, Hanxue Zhang, Xuanxu Wang, Junting Dong, Chao Gao, Mingfei Han, Tao Lu and 4 more
SERA addresses long-horizon robot planning by wrapping a fixed vision-language-action executor in an editable harness around a high-level vision-language model. Within episodes, it updates task context from observations and execution feedback; across development rollouts, agents revise planning guidance, assessment criteria, or control rules. On RoboMME, autonomous SERA reaches 51.38%, versus 44.38% without self-evolution, a 7.00-percentage-point aggregate gain, not a uniform task-level result.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.