Reproduce “Towards Agentic Robots: Model, Data, Environment, and Harness” 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
Chonghan Qin, Dongming Wu, Kaituo Feng, Jiaxu Wang, Xinlei Yu, Peiwen Sun, Yijia Fan, Xilin He and 8 more
The authors survey agentic robots as systems that use goals, observations, capabilities, and execution feedback to choose actions, gather information, revise plans, or recover. They organize prior work into four domains and seventeen overlapping categories. An external RoboDojo evaluation reports 22.48% average success across 42 simulation tasks, versus 19.34% for trained DM0.5; this comparison does not isolate feedback’s effect.
Key insights
- Agentic robotics is defined by how a system selects, coordinates, and revises capabilities during execution; it does not require complete independence from people.
- Reliability depends on the full system: stale memory, infeasible plans, weak tool interfaces, or unvalidated skill changes can undermine a capable model.
- Individual studies illustrate that harness design can affect results: one cited system’s zero-shot success rose from 60% to 82% with smaller action steps, while a different interface change reduced question-answering accuracy from 74.0% to 59.1%.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.