What happens when you give AI coding agents a lab full of robotic arms, some computing resources, and a “generous token budget” to teach the robots various tasks? The agents can apparently trace a training regimen that teaches robots to successfully cut zip ties and even insert GPUs into thin sockets on motherboards.
A glimpse of how AI could function in a completely autonomous manner to automate robot training was made possible by a new Agent Harness framework—software that wraps around AI models to enable the use of a variety of tools, as well as providing capabilities such as memory, context, constraints, and feedback loops. That agentic harness, called ENPIRE, was developed by robotics researchers at the NVIDIA GEAR (Generalist Embodied Agent Research) Lab with collaborators at Carnegie Mellon University in Pittsburgh and the University of California, Berkeley.
“A part of our NVIDIA GEAR Lab now tirelessly improves itself overnight,” Jim Fan, director of AI at NVIDIA, wrote in a LinkedIn post. “We just read the report this morning.”
Fan jokingly described the goal of such AI-guided robot training by saying, “We all celebrate holidays and Jensen doesn’t even know,” in reference to Nvidia founder and CEO Jensen Huang. But it’s not just Nvidia robotics researchers who could benefit — Fan said the team will be open-sourcing everything so anyone can host their own “self-driving robot lab at home.”
The ENPIRE harness has four modules that enable AI coding agents to perform automatic resets and validations on tasks, refine policies that guide robotic behavior, evaluate such policies across multiple physical robots working in parallel, and detect failures by analyzing logs, incorporating research papers, and improving the training infrastructure and algorithm code. More technical details are available in the research paper uploaded on June 16, 2026.
The harness was tested with three different AI coding agents, including OpenAI’s Codex with GPT-5.5, Anthropic’s Cloud Codex with Opus 4.7, and Moonshot AI’s Kimi Codex with K2.6. Teams of coding agents independently developed different algorithmic approaches to robot training, tested them in real-world experiments, and then retained any changes that helped increase the overall success rate over repeated cycles of self-guided testing.
<a href