NVIDIA ENPIRE Research Enables Autonomous Robotic Arms to Perform High Precision Assembly Jobs Without Pre Written Code
NVIDIA has unveiled a research project called ENPIRE that enables groups of robotic arms to autonomously learn and perform high precision physical assembly jobs. According to an NVIDIA research paper detailing the work, the system has zero pre written software code. Instead, the researchers utilize an ensemble of artificial intelligence agents to write code for the robots on the fly and achieve 99% success on dexterous manipulation tasks with a closed loop physical optimization approach.
Typically, programming robots for new physical tasks requires manual coding and human supervision, which can lead to the slow pace at which a robot can be adapted to new tasks. To address this issue, ENPIRE takes an automated approach to physical learning via its closed loop feedback system. This architecture has 4 modules, enabling a research endeavor to be entirely self directed. The Environment module includes automated resetting and outcome verification to standardize each trials starting condition. The Policy Improvement module automatically generates the code that will instruct the robots and improves on those commands over time. Rollout uses parallel execution on the hardware to trial control policies generated by the Policy Improvement module. Lastly, the Evolution module evaluates and prunes less effective algorithmic hypotheses based on feedback.
With this closed loop feedback structure, the research system is also capable of learning on physical hardware. For example, when it is attempting to insert a GPU into a socket, it can autonomously remove the card from a failed trial, move it to another random position on the table, and try again without human intervention to reset the scene. When it is time for a test, computer vision is utilized to evaluate if the operation was successful, and its performance metric is calculated before beginning the next iteration of the trial.
The system relies on a collection of 8 distinct coding agents each with their own running language models, including OpenAI Codex with GPT 5.5, Claude Code with Opus 4.7, and Kimi Code with Kimi K2.6. To optimize the learning process for its physical task, the agents are given a budget of language tokens, then directed to efficiently come up with a solution for the objective. Using various methods of learning from prior experiences, the AI can learn to control behavior using techniques like behavior cloning, offline reinforcement learning, or more rule based heuristic programming.
Video demonstrations show robots collaborating, with one manipulator picking a graphics card and delivering it to a second manipulator that inserts the card into a PCIe slot by aligning the contacts correctly. Other demonstrations show robots sort metallic pins into a container, use a zip tie to fasten objects, and use a cutter to snip zip ties. Success is measured with a metric that measures the pass@8 performance of the robots, meaning it takes the robot at most 8 in context tries to recover from a mistake - it indicates real adaptation, not statistical success.
Even with the project’s impressive achievements, the researchers point to some crucial bottlenecks they experienced when attempting to scale the number of robots in their fleet. The first is the rate of Mean Robot Utilization decreases, as there is a direct tradeoff between actively running hardware and running computationally intensive parts of the learning process, such as log analysis, compilation and internal code coordination among various branches of the same solution that represent different attempts. If your robots are constantly processing code logs, the hardware is not performing work.
Furthermore, as the fleet is expanded, there is a sharp rise in total token consumption, as communication among more robots results in greater computing overhead. NVIDIA developed a set of two benchmarks: Mean Robot Utilization and Mean Token Utilization to try to help the industry navigate this problem as research transitions from simulated to real world robotics applications.
