Researchers at NVIDIA have developed ENPIRE, software that enables physical robots to undergo an autonomous experimentation and execution loop similar to that used by agentic AI systems. The framework is described as a harness for coding agents and is implemented with four core modules: an Environment module (EN) for automatic reset and verification, a Policy Improvement module (PI) that triggers policy refinement, a Rollout module (R) to evaluate policies with single or multiple physical robots running in parallel, and an Evolution module (E) where agents analyze logs, consult literature, and improve training infrastructure and algorithm code to address failure modes.
How it works and why it matters
ENPIRE operates like coding agents supervising physical hardware: a scaffold controls robots assigned to complete tasks, the robots try different strategies, fail and learn, and the system both scores their performance and resets the scene when necessary. According to the authors, this closed‑loop design “transforms real‑world robot learning into a controllable optimization procedure that agents can manage, thus minimizing human effort while allowing fair ablations across training recipes and agent variants.”
Two critical components for the system are an automatic evaluation mechanism that can score each trial “without human judgement,” and an automatic reset mechanism that “returns the scene to a fresh initial state for the next trial.” The researchers note that more complex tasks may still require human effort for evaluation or resetting, which limits the complexity of tasks the system can autonomously tackle to the degree we can automate those functions.
Hardware and test setup
Each station in the experiments comprises two YAM (Yet Another Manipulator) arms from I2RT in a fixed bimanual configuration, a set of cameras, and a single workstation that runs the FastAPI server, policy inference, and the station’s agent. Each workstation is equipped with an NVIDIA RTX 5090 GPU.
Results on real‑world tasks
The team reports that frontier coding agents can autonomously develop policies that achieve up to a 99% success rate on challenging, dexterous manipulation tasks in the real world, such as PushT, organizing pins into a pin box, and using a cutter to cut a zip tie. They also tested a task in which a robot inserts GPUs into a motherboard.
Comparisons between system backbones showed tradeoffs: GPT‑5.5 within Codex and Opus 4.7 within Claude Code alternated in providing the best performance, while Kimi‑2.6 lagged behind. The researchers also observed returns to scale: larger numbers of agents (for example, eight) reached higher scoring solutions earlier, and multi‑agent setups sometimes achieved a higher absolute score than single‑agent setups, likely because they explored more of the solution space.
Scaling and instrumentation challenges
The authors highlight practical limits when scaling robot fleets. Coding agents do not fully utilize robot resources while they are reading logs, writing code, debugging, or waiting for the language‑model backbone. As the number of robots increases, robot resource utilization (MRU) falls while GPU active utilization rises. In short, there are infrastructure challenges to efficiently parallelizing many robot agents.
Conclusion
ENPIRE demonstrates a concrete approach for agent‑driven, closed‑loop policy self‑improvement on physical robots, achieving high success rates on selected manipulation tasks using YAM arms and NVIDIA RTX 5090 workstations. While the experiments show promising automation of evaluation and reset, the team cautions that task complexity and fleet scaling remain constrained by current capabilities in automatic assessment, reset procedures, and infrastructure utilization.
