NVIDIA ENPIRE: AI coding agents teach robots GPU installs

NVIDIA used a new agent harness called ENPIRE to let AI coding agents autonomously design and run robot training, producing robots that could cut zip ties and insert GPUs into motherboards. The research paper describing the system was uploaded on June 16, 2026, and the team says the approach hit a 99 percent success rate on several manipulation tasks.

How does the ENPIRE harness let AI agents train robots?

ENPIRE wraps AI models with four modules that handle automatic reset and verification, policy refinement, parallel evaluation across physical robots, and failure analysis and infrastructure improvement. The harness gives coding agents the ability to perform resets and verification on tasks, refine policies that guide robotic behavior, evaluate those policies across multiple physical robots working in parallel, and address failures by analyzing logs, ingesting research papers, and improving training infrastructure and algorithm code.

ENPIRE was developed by robotics researchers at the NVIDIA GEAR lab with collaborators at Carnegie Mellon University and the University of California, Berkeley. Jim Fan, director of AI at NVIDIA, wrote on LinkedIn, "A part of our NVIDIA GEAR lab now self-improves tirelessly overnight."

What did the experiments with AI coding agents show?

AI coding agents equipped with ENPIRE achieved high success rates across a set of manipulation tasks, reaching 99 percent success for several tasks and nearly 100 percent on pin insertion faster than a frontier human-in-the-loop method. The harness was tested with three coding agents: OpenAI's Codex with GPT-5.5, Anthropic's Claude Code with Opus 4.7, and Moonshot AI's Kimi Code with Kimi K2.6.

Experiments included the Push-T task, organizing pins in a pin box, tying and cutting zip ties, and placing a GPU into a motherboard before unplugging it to reset the trial. Larger teams of coding agents sped training: an eight-agent team hit 99 percent success on the Push-T task in two hours, a four-agent team required three hours, and a single agent took nearly five hours. The pin insertion and organization scenario produced the most promising result, where AI coding agents achieved nearly 100 percent success faster than the frontier human-in-the-loop method developed by many of the same researchers.

The team also observed limits. Robots often sat idle while agents read logs, wrote code, debugged, or waited for the language-model backbone. Larger agent teams spent more time summarizing each other’s ideas and sometimes failed to fully use available compute when launching parallel training. Higher agent counts and faster progress also increased token consumption.

Why does this matter?

ENPIRE shows an approach where language-model based coding agents can autonomously iterate on robot training at scale, reducing the need for continuous human direction. That shifts the bottleneck from generating training plans to managing compute, token costs, and orchestration of agents and robots. For labs that can supply equipment and tokens, the method promises faster policy discovery; for companies that price tokens, the higher token consumption is a concrete cost pressure.

The result is also a practical demonstration of agentic systems producing real-world robot behaviors, not just simulated plans. The team said it will open-source ENPIRE so anyone can host their own "self-running robot lab at home," a claim that ties the experimental results to a wider adoption vector.

What to watch

Watch for the ENPIRE open-source release and the accompanying research artifacts announced around the June 16, 2026 paper, which will show how accessible the harness is for other labs. Also monitor token-pricing changes from AI providers, since the experiments note higher token consumption when scaling agent teams and parallel runs.