NVIDIA's AI Agents Autonomously Train Robots Overnight

NVIDIA researchers have demonstrated a system where AI coding agents can independently train robots to perform intricate physical tasks—including inserting graphics cards into motherboard sockets and cutting zip ties—without human intervention.

The work, first reported by Ars Technica, centers on ENPIRE, a new agent harness framework developed by NVIDIA's GEAR (Generalist Embodied Agent Research) lab with collaborators from Carnegie Mellon University and UC Berkeley. The framework enables AI models to autonomously manage the entire robot training cycle, from designing experiments to analyzing failures and refining approaches.

How the system works

ENPIRE provides four core modules that handle automatic task reset and verification, policy refinement for robotic behavior, parallel evaluation across multiple physical robots, and failure analysis through log review and research paper ingestion. The system can modify training infrastructure and algorithm code based on what it learns.

Researchers tested ENPIRE with three different AI 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. These agents independently developed distinct algorithmic approaches, tested them on real hardware, and retained improvements that increased success rates across repeated cycles.

The AI-directed training achieved 99 percent success rates across several manipulation tasks, including the standard Push-T benchmark, pin organization, zip-tie operations, and GPU installation and removal.

Team size matters—with tradeoffs

Larger teams of up to eight AI coding agents reached high success rates faster than smaller configurations. An eight-agent team achieved 99 percent success on Push-T in two hours of research time, compared to three hours for a four-agent team and nearly five hours for a single agent working alone.

However, the research revealed significant limitations. Robots often remained idle while coding agents spent time reading logs, writing code, or waiting for language model responses. Larger teams consumed more time summarizing each other's ideas rather than actively using the robots, and agents sometimes failed to fully utilize available compute resources.

The faster results from larger teams also came with substantially higher token consumption—a meaningful cost consideration as AI service providers like Anthropic contemplate pricing changes that would increase token-related expenses.

Why it matters

This research represents a shift from AI as a tool that assists human robot trainers to AI as an autonomous system that can manage the entire training pipeline. For robotics labs and manufacturing facilities, the ability to run continuous improvement cycles overnight without human supervision could dramatically accelerate development timelines. The pin insertion task, where AI agents outpaced a frontier human-in-the-loop method developed by the same researchers, suggests these systems can already exceed human efficiency in specific scenarios. NVIDIA plans to open-source the framework, potentially democratizing access to autonomous robot training capabilities beyond well-funded research labs.

Jim Fan, director of AI at NVIDIA, noted that part of the GEAR lab now self-improves overnight, with researchers simply reviewing reports each morning. The team plans to open-source the complete system, allowing others to establish their own autonomous robot training labs.

The research paper with technical details was uploaded on June 16, 2026. Details were first reported by Jeremy Hsu at Ars Technica.