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, Guanya Shi
4 min
Abstract
Achieving dexterous robotic manipulation in the real world heavily relies on human supervision and algorithm engineering, which becomes a central bottleneck in the pursuit of general physical intelligence. Although emerging coding agents can generate code to automate algorithm search, their successes remain largely confined in digital environments. We conjecture that the missing abstraction to automate robotics research is a repeatable feedback loop for real-world policy improvement: reset the scene, execute a policy, verify the outcome, and refine the next iteration. To bridge this gap, we introduce ENPIRE, a harness framework for coding agents that instantiates this physical feedback routine with four core modules: an Environment module (EN) for automatic reset and verification, a Policy Improvement module (PI) that launches policy refinement, a Rollout module (R) to evaluate policies with one or multiple physical robots operating in parallel, and an Evolution module (E) in which coding agents analyze logs, consult literature, improve training infrastructure and algorithm code to address failure modes. This closed-loop system transforms real-world manipulation learning into a controllable optimization procedure, minimizing human effort while allowing fair ablations across training recipe and agent variants. Powered by ENPIRE, frontier coding agents can autonomously train a policy to achieve a 99% success rate on challenging, dexterous manipulation tasks, such as organizing a pin box, fastening a zip tie, and tool use, a process that further accelerates when we dispatch an agent team on a robot fleet. Our results suggest a practical and scalable path toward deploying coding agents to autonomously advancing robotics in the physical world.
Alex: So it's not just one robot grinding through trial and error—it's a team, each one exploring a different path, and they pool the best ideas?
Sam: Exactly. The researchers visualize this as an "idea tree." Each branch is a different strategy. The system tests them, prunes the ones that don't improve performance, and builds on the ones that do. It's an efficient way to climb toward a reliable, polished movement.
Alex: That sounds like a real time-saver. But is there a point where adding more robots makes things worse—maybe because they spend too much time coordinating with each other?
Sam: That is a critical trade-off. The researchers tracked something they call "Mean Robot Utilization"—which measures how much time each robot spends actually working versus sitting idle, waiting for instructions. As you add more robots, you get faster results, but you also consume more computing resources.
Alex: Right—and the cost isn't just electricity. These AI agents have to "think" constantly to coordinate, and that thinking uses what the researchers call "tokens," which are essentially the units of text the AI processes. More robots means more thinking, which means more cost.
Sam: Exactly. Scaling up the fleet creates diminishing returns. You might get the answer twice as fast, but you could be paying three times as much in computing overhead. It's a genuine engineering trade-off, not a free lunch.
Alex: So the real insight isn't just "robots can teach themselves"—it's that you have to budget your thinking resources just as carefully as your physical ones.
Sam: That's a good way to put it. By treating the physical world as a programmable interface, this research moves the bottleneck away from human supervision and toward how efficiently these autonomous loops can run. The future of robotics may not just be about better motors—it may be about how well we manage the intelligence coordinating them.
Alex: Thanks for listening to ResearchPod.