ENPIRE framework lets coding agents self-improve robot policies in the real world
ENPIRE is a closed-loop harness framework that enables coding agents to autonomously train and refine dexterous robot manipulation policies in the real world, achieving a 99% success rate on challenging tasks without continuous human supervision.
Score breakdown
ENPIRE removes the need for continuous human supervision and manual algorithm engineering — identified in the paper as the central bottleneck in physical robot learning — by giving coding agents a fully automated, closed-loop path to self-improve real-world manipulation policies.
- 01ENPIRE is a harness framework that gives coding agents a repeatable physical feedback loop for real-world robot policy improvement.
- 02The framework has four modules: Environment (EN), Policy Improvement (PI), Rollout (R), and Evolution (E).
- 03The Environment module handles automatic scene reset and outcome verification.
Wenli Xiao, Jia Xie, and Tonghe Zhang present ENPIRE, a harness framework designed to close the gap between coding agents' success in digital environments and the demands of real-world robotics research. The authors argue that the missing abstraction for automating robotics research is a repeatable physical feedback loop: reset the scene, execute a policy, verify the outcome, and refine the next iteration. ENPIRE instantiates this loop with four core modules — Environment (EN), Policy Improvement (PI), Rollout (R), and Evolution (E) — forming a closed-loop system that transforms real-world manipulation learning into a controllable optimization procedure.
The Evolution module is particularly notable: coding agents within it analyze execution logs, consult literature, improve training infrastructure, and modify algorithm code to address observed failure modes.
The Evolution module is particularly notable: coding agents within it analyze execution logs, consult literature, improve training infrastructure, and modify algorithm code to address observed failure modes. This design minimizes human effort while enabling fair ablations across training recipes and agent variants. Deployed on challenging dexterous manipulation tasks — organizing a pin box, fastening a zip tie, and tool use — ENPIRE-powered frontier coding agents achieved a 99% success rate autonomously. The paper also reports that performance accelerates further when an agent team is dispatched across a robot fleet operating in parallel, suggesting a scalable path toward coding agents that can autonomously advance robotics in the physical world.
Key facts
- 01ENPIRE is a harness framework that gives coding agents a repeatable physical feedback loop for real-world robot policy improvement.
- 02The framework has four modules: Environment (EN), Policy Improvement (PI), Rollout (R), and Evolution (E).
- 03The Environment module handles automatic scene reset and outcome verification.
- 04The Evolution module has coding agents analyze logs, consult literature, and improve training infrastructure and algorithm code.
- 05Frontier coding agents powered by ENPIRE achieved a 99% success rate on dexterous manipulation tasks.
- 06Tasks demonstrated include organizing a pin box, fastening a zip tie, and tool use.
- 07Performance accelerates further when an agent team is dispatched across a robot fleet operating in parallel.
Topics
Summary and scoring are generated automatically from the original article. We always link back to the publisher and never republish images or paywalled content. Last processed Jun 19, 2026 · 10:25 UTC. How this works →