Nvidia researchers handed AI coding agents a lab full of robotic arms and a token budget, and the agents taught the robots to insert GPUs into motherboard sockets, cut zip ties, and organize pins — hitting a 99% success rate across manipulation tasks. The framework, called ENPIRE, was built by Nvidia's GEAR lab with collaborators at Carnegie Mellon University and the University of California, Berkeley, and described in a paper uploaded June 16, 2026. The team plans to open-source the entire stack.
ENPIRE is an agent harness — software that wraps around a language model to give it tools, memory, context windows, constraints, and feedback loops. Its four modules let coding agents automatically reset and verify physical tasks, refine the policies that drive robot behavior, run parallel evaluations across multiple robots, and diagnose failures by reading logs, ingesting research papers, and rewriting their own training code.
The researchers tested ENPIRE with three coding agents: OpenAI's Codex running on GPT-5.5, Anthropic's Claude Code on Opus 4.7, and Moonshot AI's Kimi Code on Kimi K2.6. Each team independently devised algorithmic approaches to robot training, ran them on real hardware, and kept whatever changes raised the success rate over repeated self-directed cycles.
Key facts
- 01Nvidia's ENPIRE harness let AI coding agents hit a 99% success rate across manipulation tasks including GPU insertion and zip-tie cutting.
- 02An 8-agent team solved the standard Push-T task in 2 hours, versus 3 hours for a 4-agent team and nearly 5 hours for a single agent.
- 03ENPIRE was tested with OpenAI's Codex on GPT-5.5, Anthropic's Claude Code on Opus 4.7, and Moonshot AI's Kimi Code on Kimi K2.6.
- 04The research paper was uploaded June 16, 2026 by Nvidia GEAR, Carnegie Mellon University and UC Berkeley, with plans to open-source the framework.
- 05On pin insertion and organization, the agents reached nearly 100% success faster than a frontier human-in-the-loop method from the same researchers.
Team size mattered. On the standard Push-T benchmark — sliding a T-shaped block into a target position on a table — an 8-agent team reached 99% success in 2 hours of research time. A 4-agent team needed 3 hours. A single agent took nearly 5 hours. The scaling was clean enough that the researchers think bigger swarms of coding agents may be the right unit of work for autonomous robotics R&D, not solo models.
The most striking result came on a pin insertion and organization task, where the agents hit nearly 100% success faster than a frontier human-in-the-loop method developed by some of the same researchers. That is the rare case of an autonomous agent loop beating its own designers at their day job, on their own benchmark.
Jim Fan, director of AI at Nvidia, framed the work on LinkedIn as a piece of the lab that now runs without supervision, with researchers reading the morning reports rather than babysitting training runs. He also said the team would release everything so outside labs — and hobbyists — could host a self-running robot lab at home.
“We all take a holiday and Jensen wouldn't even notice.”— Jim Fan, Director of AI at Nvidia
The limits showed up in the seams. Robots often sat idle while coding agents read logs, wrote code, debugged, or waited on the language-model backbone. Larger teams spent more time summarizing each other's ideas and less time actually moving the robots, and the agents sometimes failed to saturate available compute when launching parallel training sessions. Faster results also came with sharply higher token consumption — a real cost line at a moment when Anthropic and others are weighing price changes that would push token bills up.
ENPIRE lands in the middle of a broader Nvidia push into physical AI. On May 31, Nvidia announced a partnership with Chinese robotics firm Unitree to supply a Reference Humanoid Robot for research labs. In early June, Jensen Huang toured South Korea and met Hyundai Motor Executive Chair Chung Euisun to discuss mass-manufacturing AI-powered robots. Hyundai owns Boston Dynamics, which sells the Spot quadruped and is commercializing the Atlas humanoid. Last week's Nvidia–LG Group AI factory announcement, which AI Chat Daily covered, fits the same pattern: Nvidia is bundling chips, software, and now training methodology into a full-stack robotics offering.
The interesting wrinkle is what ENPIRE implies about the bottleneck in robotics. For a decade, the limiting reagent has been engineer-hours: humans writing reward functions, tuning policies, resetting hardware, debugging perception stacks. If 8 coding agents on a generous token budget can compress that work into a single overnight cycle and beat a human-in-the-loop baseline, then the cost curve for new robot skills starts to look a lot more like the cost curve for software — driven by tokens, not headcount.
Open-sourcing the harness is the move that matters for the field. A closed ENPIRE would consolidate robotics research inside Nvidia's lab. An open one turns every academic group and well-funded startup into a potential self-improving robot shop, and routes the dependency to whichever coding agent — Codex, Claude Code, Kimi Code — proves most reliable per dollar. That is a quietly enormous demand pull for frontier coding models, and a fresh axis of competition between OpenAI, Anthropic, and Moonshot AI that has nothing to do with chatbot benchmarks.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
