Nvidia published on June 17, 2026 the ENPIRE framework, in which teams of AI coding agents — using models from OpenAI, Anthropic, and Moonshot — independently train robotic arms for tasks such as GPU installation and zip-tie cutting. The study, conducted jointly with Carnegie Mellon University and UC Berkeley, shows that eight agents achieve 99 percent success rates in two hours of work without human intervention.
Key takeaways
- ENPIRE wraps AI models in a harness enabling autonomous reset, verification, and refinement of robot motion policies
- Eight AI agents completed the Push-T task with 99% success in 2 hours — four times faster than a single agent
- Agents approached 100% success on a pin insertion task faster than a reference human-in-the-loop method
- Code and research paper available since June 16, 2026
- Jim Fan, AI director at Nvidia GEAR lab: "Part of our NVIDIA GEAR lab now self-improves tirelessly overnight. We just read the reports in the morning."
How ENPIRE works
ENPIRE is not an AI model but a harness — software that wraps coding agents in a structured working environment with robots. The framework developed by the NVIDIA GEAR lab consists of four modules operating in a loop: task reset and verification, policy refinement, parallel evaluation across multiple physical robots, and failure handling through log analysis and research paper ingestion.
Results and limitations
Tests were conducted with three AI coding agents: Codex (GPT-5.5, OpenAI), Claude Code (Opus 4.7, Anthropic), and Kimi K2.6 (Moonshot AI). Eight agents achieved 99% success on the Push-T task in two hours. Four agents needed three hours, and a single agent required nearly five.
On the pin insertion task, eight agents approached 100% success faster than a reference method with human involvement at each step. Researchers identified three limitations: larger teams spent more time summarizing each other's activities; robots often sat idle while agents wrote code; agents did not always effectively manage compute resources when launching parallel training sessions.
Nvidia GEAR lab context
In May 2026, the same team announced a partnership with Unitree Robotics for the Isaac GR00T Reference Humanoid Robot — a platform combining Unitree H2 Plus hardware with Jetson Thor and Sharpa Wave tactile hands. Jensen Huang met with Hyundai Motor Group leadership in South Korea to discuss mass manufacturing of AI-powered robots — Hyundai owns Boston Dynamics.
ENPIRE fits into Nvidia's broader strategy: not just supplying GPUs, but building a complete software stack for Embodied AI — from simulators (NVIDIA Isaac Lab, Cosmos) through training frameworks to agentic tools.
Why this matters
Autonomous robot training by AI agents eliminates one of the most costly and time-consuming elements in designing manipulation systems: the human-in-the-loop feedback cycle. Traditionally, a researcher writes policy code, runs the robot, observes results, corrects, and repeats. ENPIRE delegates this cycle to agents, which do it faster, without breaks, and with the ability to test in parallel across multiple robots simultaneously. The 99% success rate on pin insertion — a task requiring sub-millimeter precision — in less time than a human-in-the-loop approach is concrete evidence that agent autonomy in a robotics lab is no longer a research concept. This remains early academic work with constrained tasks in a controlled environment, but the direction is clear.
What's next
- Nvidia announced an open-source release of the full ENPIRE framework — anyone will be able to run their own "self-running robot lab at home"
- The research paper published June 16, 2026 contains full technical specifications and experimental configurations
- The next step for the GEAR team is to extend tests to more complex tasks outside laboratory environments
Sources
- Ars Technica — AI coding agents taught robots how to install GPUs and cut zip ties
- Nvidia GEAR Lab — ENPIRE project page
