ENPIRE agentic robot self-improvement, 99% success rate

ENPIRE, introduced in a paper submitted 18 Jun 2026 by Wenli Xiao and 16 coauthors, is a harness framework that makes real-world robot policy improvement repeatable and automatable. The system instantiates a closed-loop physical feedback routine and, powered by coding agents, can autonomously train a policy to achieve a 99% success rate on dexterous manipulation tasks such as organizing a pin box, fastening a zip tie, and tool use.

What is ENPIRE?

ENPIRE is a four-module framework that turns real-world manipulation learning into a controllable optimization procedure: Environment (EN), Policy Improvement (PI), Rollout (R), and Evolution (E). The paper frames the cycle as reset the scene, execute a policy, verify the outcome, and refine the next iteration, chaining those steps so coding agents can iteratively address failure modes and reduce human supervision.

The authors present ENPIRE as a "harness framework for coding agents" that minimizes human effort while enabling fair ablations across training recipes and agent variants. The submission emphasizes that the framework allows multiple physical robots to operate in parallel during evaluation.

How does the system work in practice?

ENPIRE runs a repeatable feedback loop with four core modules: Environment for automatic reset and verification, Policy Improvement to launch policy refinement, Rollout to evaluate policies with one or multiple physical robots operating in parallel, and Evolution where coding agents analyze logs, consult literature, and improve training infrastructure and algorithm code. Each module has a specific role in the loop: EN resets and checks outcomes, R executes and gathers rollouts, PI updates the policy, and E drives higher-level changes to address systematic failures.

Practically, the paper shows this closed-loop approach lets frontier coding agents autonomously train policies to a reported 99% success rate on challenging dexterous tasks. The authors note the process further accelerates when they "dispatch an agent team on a robot fleet," indicating the framework supports parallelized experimentation across multiple robots.

The submission positions ENPIRE against the prevailing bottleneck in dexterous robotic manipulation, which the authors identify as reliance on human supervision and algorithm engineering. By automating the reset-evaluate-refine cycle and supporting coding agents that can modify code and infrastructure, ENPIRE aims to close that gap between success in digital coding-agent settings and the physical world.

Why it matters

ENPIRE constrains an open-ended research problem into a repeatable optimization loop, which changes what can be automated: not only policy tuning but also algorithm and infrastructure fixes driven by coding agents. Achieving a 99% success rate on real dexterous tasks implies the framework can convert some forms of hands-on engineering work into iterative agent-led development. Dispatching agent teams across robot fleets suggests a path to scale experiments without linearly increasing human supervision.

What to watch

Look for replication and code or dataset releases tied to this arXiv submission, and for follow-up experiments that measure how much faster training proceeds when an agent team runs on a robot fleet versus a single robot. Confirming that the 99% success rate generalizes beyond the listed tasks would be the clearest sign ENPIRE scales beyond the paper's demonstrations.

Authors credited on the submission include Wenli Xiao, Jia Xie, Tonghe Zhang, Haotian Lin, Letian "Max" Fu, Haoru Xue, Jalen Lu, Yi Yang, Cunxi Dai, Zi Wang, Jimmy Wu, Guanzhi Wang, S. Shankar Sastry, Ken Goldberg, Linxi "Jim" Fan, Yuke Zhu, and Guanya Shi. The paper is archived on arXiv as arXiv:2606.19980 (cs.AI), submitted 18 Jun 2026.