ENPIRE: Agentic Robot Policy Self-Improvement in the Real World

Closing the Loop Between Code and Contact

Achieving dexterous robotic manipulation in the real world currently relies heavily on human supervision. Constant algorithmic engineering creates a massive bottleneck in the pursuit of general physical intelligence. While AI coding agents are proficient at automating research in digital environments, they struggle with unpredictable physics.

Researchers have developed a system called ENPIRE. This system lets AI coding agents act like robotic scientists. Instead of a human tweaking parameters, these agents set up robot environments. They run experiments, analyze logs, and rewrite their own code to improve performance. The system transforms robot learning into a controllable optimization procedure.

The bottleneck of manual dexterity

The status quo in robotics research is a cycle of human-intensive trial and error. To teach a robot a delicate task, engineers must manually collect data. They must also design reward functions (mathematical signals that tell the robot if it is succeeding). Finally, they must intervene to reset the scene when a task fails.

Even with advanced methods like Reinforcement Learning (RL; a way for agents to learn through trial and error), human labor is a limiting factor. Humans must "babysit" the process to manage data collection and algorithm adjustments.

Current coding agents can navigate software repositories and fix bugs. However, they lack a meaningful interface with the physical world. They cannot "feel" a failed grasp or "see" a misalignment without human translation. Without an automated reset and verification loop, the jump to physical policy improvement remains difficult.

A four-stage engine for physical research

The authors introduce ENPIRE to bridge this gap. The framework treats real-world manipulation as a repeatable feedback loop. As shown in, the system uses four modules: Environment (EN), Policy Improvement (PI), Rollout (R), and Evolution (E).

The process begins with a human-guided stage. Coding agents construct an autonomous environment here. They use procedural tool calls to build immutable APIs (standardized sets of commands). These APIs handle safety constraints, automated verification, and automated resets.

Once these APIs are ready, the system enters a fully autonomous second stage. In this phase, the Policy Improvement module launches the actual research. The coding agent has write access to a training codebase. It analyzes logs and consults literature to modify the training code.

To scale this, the Evolution module allows a decentralized team of agents to work across a robot fleet. These agents use Git to collaborate. They share successful "recipes" and discard failing hypotheses. This allows them to optimize code search without human intervention.

Scaling success through robot fleets

The authors demonstrate that ENPIRE can "hill-climb" success rates on complex tasks. In the pin insertion task, the paper reports that the agent achieves a 100% success rate. This happens faster than frontier human-in-the-loop methods.

The system is also scalable. By increasing the number of parallel agent-robot pairs, researchers reduced the wall-clock time (actual elapsed time) required for high performance. For the Push-T task, scaling from one agent to eight reduced the time to reach a perfect score from five hours to two hours .

In pin insertion, the time to reach near-perfect success dropped from 1.5 hours to 40 minutes when using eight robots.

The authors propose two new metrics to track efficiency: Mean Robot Utilization (MRU) and Mean Token Utilization (MTU). MRU measures how much time the robot spends actively executing experiments. MTU measures how many tokens (units of text processed by an AI) the fleet consumes.

The costs of autonomy

The authors identify clear trade-offs in the ENPIRE architecture. One limitation is the underutilization of resources. Coding agents spend time reading logs, writing code, or waiting for the Large Language Model (LLM) to respond. Consequently, the robots often sit idle. The paper finds that as the fleet size increases, MRU actually decreases [Figure 7a].

There is also a steep economic penalty for speed. While a larger fleet reduces the time to success, token costs grow super-linearly. This means costs rise much faster than the number of agents added. The authors report that the total token budget required for success increases more rapidly than the reduction in wall-clock time [Figure 7c]. Larger fleets reach success sooner, but they require a disproportionately higher token budget.

A verdict on autonomous labs

ENPIRE represents a functional prototype for autonomous robotics research. It moves the research process from the human engineer to the coding agent. Achieving a 99% success rate on tasks like zip-tie cutting suggests the loop is viable.

If you want to deploy this today, consider your budget. The system can accelerate research if you have the computational resources for high token costs. It is not a "set it and forget it" solution for low-resource labs. However, it provides a blueprint for turning a robot fleet into a self-improving engine. Code and demonstrations are reportedly available at research.nvidia.com/labs/gear/enpire.

Figures from the paper