Version 2

Two systems for autonomous robot training appeared in the same week in June 2026. NVIDIA's ENPIRE framework, developed with Carnegie Mellon University and UC Berkeley, uses AI coding agents to run robot training experiments overnight without human supervision, demonstrated on dexterous manipulation tasks including GPU installation [1]. Anthropic's Project Fetch Phase 2 showed Claude Opus 4.7 independently programming an unfamiliar robot dog in 12 minutes and 7 seconds — a task that took a human-assisted team roughly 4 hours in 2024 [5][4][3]. Both systems operate against real hardware. The Project Fetch result attracted broad social amplification in the 24 hours following its release [6][7][8][9][10][11].

If AI agents can autonomously close the trial-and-error loop in robot training — writing code, reading sensor data, adjusting policies, and repeating — robots can acquire new skills without proportional human labor. Both NVIDIA and Anthropic demonstrated this on real hardware, not simulation, in the same week.

NVIDIA's GEAR lab, working with Carnegie Mellon University and UC Berkeley, released ENPIRE — a framework that wraps AI coding agents with memory, context management, constraint handling, and feedback loops to run robot training experiments autonomously [1]. Researchers set tasks, then let the system operate overnight: by morning, a report describes what the agents tried and how robot performance changed. The system has been used to train robotic arms on dexterous manipulation tasks including cutting zip ties and inserting GPUs into motherboard sockets. NVIDIA's own characterization: 'A part of our NVIDIA GEAR lab now self-improves tirelessly overnight. We just read the reports in the morning' [1]. Decrypt.co covered the NVIDIA result as an example of robots that 'train themselves' using AI coding agents [2].

On the same day ENPIRE drew broad attention, Anthropic published results from Project Fetch Phase 2 [3]. In the original 2024 experiment, human Anthropic employees aided by Claude attempted to program an off-the-shelf robot dog from scratch, taking roughly 4 hours [4]. In Phase 2, Claude Opus 4.7 was given the task alone: connect real robot hardware, read camera and lidar sensor feeds, write movement code, and track the robot's location. The model completed the full sequence in 12 minutes and 7 seconds [5][4] — approximately 20x faster than the human-assisted team. The result spread quickly across social media, with multiple accounts amplifying Wes Roth's summary of the experiment [6][7][8][9][10][11].

A parallel body of academic work covers the methods underlying both systems: language-instructed skill acquisition, continual robot learning, and LLM-guided reinforcement learning for real-world manipulation [12][13][14][15]. These papers situate ENPIRE and Project Fetch within wider research on using LLMs to reduce human effort in robot learning pipelines. The NVIDIA and Anthropic work is distinctive for running on real hardware and framing results as production capabilities rather than ablation studies.

One notable gap in coverage: a social media post asserted that every major AI lab except OpenAI and Anthropic is building physical AI [16]. Project Fetch Phase 2 directly contradicts the Anthropic half of this claim, but neither company has announced a sustained robotics research division. Whether these experiments represent ongoing strategic bets or periodic capability probes remains unaddressed in available sources.