Source note
ENPIRE: Agentic Robot Policy Self-Improvement in the Real World
Real World Robot LearningRobot Policy Self ImprovementDexterous ManipulationCoding AgentsRobot Data Scaling
Summary
ENPIRE lets coding agents improve robot manipulation policies on real hardware by running reset, rollout, verification, and code-edit loops with little human work after setup. It reports up to 99% real-world success on dexterous tasks and faster improvement when eight robot-agent workers run in parallel.
Problem
- Real robot policy learning still needs humans to reset scenes, judge outcomes, collect rollouts, and change training code, which slows data scaling for dexterous manipulation.
- Coding agents can improve software in digital settings, but real robots need safe execution bounds, automatic resets, and reliable success checks before agents can run experiments.
- The problem matters because robot time and human operator time become bottlenecks when training generalist policies on contact-rich tasks.
Approach
- ENPIRE gives a coding agent task APIs for hard safety limits, automatic verification, automatic reset, rollout execution, and training-code edits.
- In the first stage, the agent builds reset and reward/checker code from short human feedback and demos; after approval, these APIs are fixed for later runs.
- In the second stage, the agent reads logs and videos, edits BC/RL/heuristic/code-policy training code, launches real robot rollouts, and keeps changes that raise measured success.
- For fleets, ENPIRE assigns one agent per robot; agents test different code branches asynchronously and share or copy successful recipes through Git.
- The paper adds Mean Robot Utilization, GPU utilization, Mean Token Utilization, tokens-to-success, and time-to-success to measure how much robot time and model-token budget the agents spend.
Results
- The abstract claims coding agents reached 99% success on real dexterous tasks including PushT, pin-box insertion, and zip-tie cutting.
- Pin insertion uses 4 mm holes and requires 50 consecutive real-world successes in evaluation; the paper reports convergence to 100% faster than a cited human-in-the-loop method.
- In simulation on Gym-PushT, Claude Code and Codex reached 95% success in about 2 hours; Kimi Code took about twice as long.
- Fleet scaling cut Push-T time to a 1.0 normalized score from about 5 hours with 1 agent to about 2 hours with 8 agents.
- Fleet scaling cut pin-insertion time to near-perfect success from more than 1.5 hours with 1 agent to about 40 minutes with 8 agents.
- The zip-tie checker was optimized to under 150 ms latency, and RoboCasa365 results are reported as better than GR00T VLA and CaP-X, but the excerpt gives no exact success rates for that comparison.