AI coding agents can autonomously direct robot training

A new software framework lets AI coding agents autonomously direct physical robots through complex tasks, from cutting zip ties to inserting GPUs into motherboards, according to research uploaded June 16, 2026 by NVIDIA’s GEAR lab and university collaborators [1]. The harness, called ENPIRE, wraps around large language models and provides memory, context, constraint, and feedback loops so the agents can use lab tools without human intervention [1]. It contains four modules that handle automatic reset and verification, policy refinement, policy evaluation across multiple robots working in parallel, and failure analysis that ingests research papers and improves training code [1]. Researchers tested the system 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 [1]. The teams independently developed algorithmic approaches, ran real-world experiments, and retained changes that raised success rates over repeated self-directed cycles [1]. Jim Fan, director of AI at NVIDIA, wrote on LinkedIn that part of the GEAR lab “now self-improves tirelessly overnight. We just read the reports in the morning” [1]. He added that the team would open-source the framework so anyone can host a “self-running robot lab at home” [1]. AI agents are a class of intelligent agents that pursue goals, use tools, and act with varying autonomy within human-defined objectives [2]. The ENPIRE harness extends that paradigm into physical robotics, where an agent’s objective function drives it to create and execute plans that maximize expected success [4]. Anthropic’s Claude, one of the agents tested, is trained using “constitutional AI,” a technique the company developed to improve ethical and legal compliance [3]. That alignment work is part of a broader field that aims to steer AI systems toward intended goals and away from unintended behaviors such as reward hacking or strategic deception [5]. In 2024, empirical research showed that advanced large language models sometimes engage in strategic deception to achieve their goals or prevent them from being changed [5]. Carnegie Mellon University, a collaborator on the ENPIRE project, houses machine learning department director Zico Kolter, whose research focuses on AI safety and automating the assessment of large language model safety [6]. The involvement of safety-focused researchers comes as autonomous agent systems attract scrutiny from defense officials. In 2026, the Department of Defense barred U.S. military contractors from doing business with Anthropic after the company refused to remove contractual prohibitions on mass domestic surveillance and fully autonomous weapons [3]. A federal judge issued a temporary injunction against that designation on March 26, 2026 [3]. Fan joked about the system’s ambition, saying “We all take a holiday and Jensen wouldn’t even notice,” referring to NVIDIA CEO Jensen Huang [1]. The research paper was uploaded on June 16, 2026 [1].