The Information Machine

AI Coding Agents Autonomously Program and Train Physical Robots Without Human Supervision · history

Version 5

2026-06-23 18:27 UTC · 71 items

What

Two autonomous robot training systems drew sustained attention across the week of June 17–23, 2026: NVIDIA's ENPIRE framework, which runs AI coding agents across 8 parallel robot stations overnight without human supervision [2][1], and Anthropic's Project Fetch Phase 2, in which Claude Opus 4.7 programmed an off-the-shelf robot dog in 12 minutes and 7 seconds — roughly 20x faster than a human-assisted team took in 2024 [6][4]. A secondary framing has emerged arguing that both systems represent a broader industry shift: LLMs writing code to control robots, rather than directly issuing motor commands [7]. Amplification continues but is producing few new substantive claims.

Why it matters

Both systems close the trial-and-error loop in robot control on real hardware without human checkpoints. The Project Fetch result is particularly notable because Claude had no robotics-specific training, suggesting general-purpose reasoning is sufficient for hardware integration tasks that previously required specialized human effort. A growing secondary argument frames this not as a robotics result but as evidence about the breadth of frontier model capability.

Open questions

  • One amplifier reports Project Fetch timing as 9 minutes / 6 hours [13] while primary sources establish 12 minutes 7 seconds / ~4 hours [6][4] — which figures are correct?

  • How well do ENPIRE-trained policies generalize across hardware configurations not seen during overnight autonomous training runs? [2]

  • Neither NVIDIA nor Anthropic has announced a sustained robotics research division — do these experiments represent ongoing strategic investment or periodic isolated probes?

  • Vivek Kotecha argues the industry spent three years on the wrong assumption (LLMs as direct robot controllers) [7] — does the success of the code-generation approach invalidate prior direct-control research, or do both approaches remain viable in different settings?

Narrative

NVIDIA's GEAR lab, in collaboration with Carnegie Mellon University and UC Berkeley, released ENPIRE — a framework that deploys AI coding agents across 8 parallel robot stations overnight [1]. Each agent writes its own reward functions, edits training code, and adjusts policies based on sensor feedback without human supervision [2][1]. Researchers set tasks, then review a morning report on what the agents tried and how robot performance changed. The system has been demonstrated on dexterous manipulation tasks including cutting zip ties and inserting GPUs into motherboard sockets. NVIDIA's framing: 'A part of our NVIDIA GEAR lab now self-improves tirelessly overnight. We just read the reports in the morning' [2].

In the same week, Anthropic published results from Project Fetch Phase 2 [3]. In a 2024 baseline, human Anthropic employees aided by Claude spent roughly 4 hours programming an off-the-shelf robot dog from scratch [4]. In Phase 2, Claude Opus 4.7 — a model with no robotics-specific training [5] — was given the task alone: connect real hardware, read camera and lidar feeds, write movement code, and track the robot's location. The model completed the full sequence in 12 minutes and 7 seconds [6], approximately 20x faster than the human-assisted team.

A growing secondary argument frames what these systems represent methodologically. Several commentators argue that both ENPIRE and Project Fetch work by having LLMs write code that controls robots — not by having LLMs issue motor commands directly [7]. On this reading, the industry spent years on the wrong assumption about how LLMs would integrate into robotics. Social reframers including ninzaverse and 0x_codex go further, arguing Project Fetch's significance is not the robot dog task but the demonstration that a general-purpose model can handle hardware integration, sensor interpretation, and code generation with no domain-specific preparation [8][5].

Concurrent academic work on language-instructed skill acquisition, continual robot learning, and LLM-guided reinforcement learning provides methodological grounding for both systems [9][10][11][12]. ENPIRE and Project Fetch are distinctive for demonstrating these methods on real hardware and presenting results as production capabilities rather than research ablations.

Timeline

  • 2024: Anthropic Project Fetch Phase 1: human employees aided by Claude program an off-the-shelf robot dog in roughly 4 hours, establishing the comparison baseline. [3][30][4]
  • 2026-06-17: Ars Technica reports on NVIDIA ENPIRE: AI coding agents autonomously train robotic arms overnight on dexterous tasks including GPU installation, in a collaboration between NVIDIA GEAR, CMU, and UC Berkeley. [2]
  • 2026-06-18: Anthropic releases Project Fetch Phase 2: Claude Opus 4.7 programs a robot dog in 12 minutes 7 seconds without human assistance, approximately 20x faster than the 2024 human-assisted effort. [6][3][31][32][4]
  • 2026-06-19: Project Fetch Phase 2 spreads on social media; Decrypt.co publishes additional coverage of NVIDIA ENPIRE. [21][22][23][24][25][20][33]
  • 2026-06-20: Reframing voices argue Project Fetch demonstrates broad autonomous capability rather than a narrow robotics demo; ENPIRE detail surfaces that 8 stations run in parallel with agents writing their own reward functions. [8][5][26][13][1]
  • 2026-06-21: Continued amplification via retweets and video platforms, with no new substantive claims. [27][28]
  • 2026-06-23: Further amplification continues; Vivek Kotecha argues the industry spent three years on the wrong assumption that LLMs would directly control robots rather than write code for them. [18][19][34][35][7]

Perspectives

NVIDIA GEAR Lab

ENPIRE enables a self-improving research lab where 8 parallel agent-driven robot stations operate overnight and researchers review reports in the morning.

Evolution: Consistent; ENPIRE is the public research instantiation of NVIDIA's broader push into physical AI infrastructure.

[2][14][1]

Anthropic

Project Fetch Phase 2 shows a frontier LLM with no robotics-specific training can independently handle hardware integration, sensor reading, code writing, and navigation far faster than a human-assisted team.

Evolution: Phase 2 directly follows Phase 1, showing expanded autonomous capability by removing the human from the loop entirely.

[3][15][16][6][17][18]

Social reframers (0x_codex, ninzaverse)

Project Fetch is evidence of general-purpose autonomous capability in physical systems, not a robotics trick — the significance is that Claude had no robotics training yet completed the task.

Evolution: Ninzaverse has now published a Medium article expanding this argument [19], reinforcing the framing beyond social posts.

[8][5][19]

Vivek Kotecha

The industry spent three years assuming LLMs would directly control robots; ENPIRE and Project Fetch demonstrate that the winning approach is LLMs writing code that controls robots.

Evolution: New voice this pass; frames both systems as a correction to a long-held industry assumption rather than an incremental advance.

[7]

Wes Roth and social amplifiers

Project Fetch Phase 2 is a noteworthy demonstration that Claude can independently program unfamiliar robot hardware; shared widely without notable skepticism.

Evolution: Continued amplification; some amplifiers report slightly different figures (9 minutes, 6 hours) than primary sources.

[20][21][22][23][24][25][26][13][27][28]

Jeremy Hsu / Ars Technica

Reports ENPIRE as a significant step toward fully autonomous robot skill acquisition pipelines, amplifying NVIDIA's 'self-improving lab' framing without skepticism.

Evolution: Consistent neutral-to-positive technology reporting.

[2]

Academic research community (CMU, UC Berkeley, USC RASC, AAAI)

Concurrent work confirms LLMs can guide robot skill acquisition in unfamiliar environments, providing methodological grounding for what ENPIRE and Project Fetch demonstrate on real hardware.

Evolution: Ongoing; papers predate or run parallel to the industry announcements.

[9][10][11][12]

Social commentator (thehype.)

Argued every major AI lab except OpenAI and Anthropic is investing in physical AI, positioning both as absent from embodied AI development.

Evolution: Claim was contradicted the same day by Project Fetch Phase 2; the original poster has not acknowledged the contradiction.

[29]

Tensions

  • Vivek Kotecha argues the winning approach is LLMs writing code that controls robots [7], implying three years of direct-control research was misdirected; this framing is not shared by NVIDIA or Anthropic, who present their systems as demonstrations rather than corrections to prior approaches. [7][2][3]
  • The claim that Anthropic is absent from physical AI development [29] is directly contradicted by Project Fetch Phase 2 [3][6]; neither company has announced a sustained robotics division, leaving open whether these experiments are ongoing strategic investment or periodic isolated probes. [29][3][6]
  • Social amplifiers report Project Fetch timing as 9 minutes / 6 hours [13] while primary sources establish 12 minutes 7 seconds / ~4 hours [6][4]; no source has addressed the discrepancy. [13][6][4]

Sources

  1. [1] Nvidia ENPIRE: 8 robot stations, each running its own AI coding agent. The agents write their own reward functions, edit... — reactive:ai-coding-agents-robot-training (2026-06-20)
  2. [2] AI coding agents taught robots how to install GPUs and cut zip ties — Ars Technica AI (2026-06-17)
  3. [3] Project Fetch: Can Claude train a robot dog? \ Anthropic — reactive:ai-coding-agents-robot-training
  4. [4] Claude Opus 4.7 programmed a robot dog from scratch in 12 minutes and 7 seconds. A human-assisted team needed roughly 4 ... — reactive:ai-coding-agents-robot-training (2026-06-19)
  5. [5] 🚨 Anthropic just had an AI operate a robot dog with zero human help. and it was never trained on robotics. — reactive:ai-coding-agents-robot-training (2026-06-20)
  6. [6] Anthropic just showed Claude Opus 4.7 program a robodog in 12:07 mint, about 20x faster than last year’s Claude-aided hu… — Rohan Paul Twitter (2026-06-18)
  7. [7] For three years, the AI industry assumed that large language models would control robots. — reactive:ai-coding-agents-robot-training (2026-06-23)
  8. [8] Anthropic’s Project Fetch is not really about a robot dog doing tricks. — reactive:ai-coding-agents-robot-training (2026-06-21)
  9. [9] Continual Robot Learning via Language-Guided Skill Acquisition | OpenReview — reactive:ai-coding-agents-robot-training
  10. [10] LLMs can help robots learn new tasks in unfamiliar places – Robotics and Autonomous Systems Center — reactive:ai-coding-agents-robot-training
  11. [11] Towards Autonomous Reinforcement Learning for Real-World Robotic Manipulation with Large Language Models — reactive:ai-coding-agents-robot-training
  12. [12] [PDF] Efficient Language-instructed Skill Acquisition via Reward-Policy Co ... — reactive:ai-coding-agents-robot-training
  13. [13] 🚨It Took Humans 6 Hours. Claude Did It Alone in 9 Minutes. — reactive:ai-coding-agents-robot-training (2026-06-20)
  14. [14] ENPIRE: Agentic Robot Policy Self-Improvement in the Real World — reactive:ai-coding-agents-robot-training
  15. [15] Read the full write-up of Project Fetch: — reactive:ai-coding-agents-robot-training
  16. [16] Anthropic's Project Fetch: How AI models like Claude can control robots | Anthropic posted on the topic | LinkedIn — reactive:ai-coding-agents-robot-training
  17. [17] Project Fetch: Phase two - Anthropic — reactive:ai-coding-agents-robot-training
  18. [18] Watch the robodogs in action in our first Project Fetch experiment: — reactive:ai-coding-agents-robot-training
  19. [19] Anthropic Ran Project Fetch Again, and This Time AI Didn't Need Us — reactive:ai-coding-agents-robot-training
  20. [20] Anthropic released Phase 2 of Project Fetch, testing whether Claude could independently program an unfamiliar robot dog. — reactive:ai-coding-agents-robot-training (2026-06-19)
  21. [21] RT @WesRoth: Anthropic released Phase 2 of Project Fetch, testing whether Claude could independently program an unfamili... — reactive:ai-coding-agents-robot-training (2026-06-19)
  22. [22] RT @WesRoth: Anthropic released Phase 2 of Project Fetch, testing whether Claude could independently program an unfamili... — reactive:ai-coding-agents-robot-training (2026-06-19)
  23. [23] RT @WesRoth: Anthropic released Phase 2 of Project Fetch, testing whether Claude could independently program an unfamili... — reactive:ai-coding-agents-robot-training (2026-06-19)
  24. [24] RT @WesRoth: Anthropic released Phase 2 of Project Fetch, testing whether Claude could independently program an unfamili... — reactive:ai-coding-agents-robot-training (2026-06-19)
  25. [25] Project Fetch Phase 2: Anthropic let Claude Opus 4.7 run a robot dog solo. — reactive:ai-coding-agents-robot-training (2026-06-19)
  26. [26] Claude outperformed humans in controlling a robot dog 🤖 — reactive:ai-coding-agents-robot-training (2026-06-20)
  27. [27] RT @WesRoth: Anthropic released Phase 2 of Project Fetch, testing whether Claude could independently program an unfamili... — reactive:ai-coding-agents-robot-training (2026-06-21)
  28. [28] NVIDIA's AI agents taught robots to seat GPUs overnight with zero human steering #AI — reactive:ai-coding-agents-robot-training
  29. [29] every big ai lab is now building physical ai. except openai and anthropic. why? — reactive:ai-coding-agents-robot-training (2026-06-16)
  30. [30] Anthropic reran Project Fetch from 2024, their robodog experiment where random employees tried to make an off the shelf,... — reactive:ai-coding-agents-robot-training (2026-06-18)
  31. [31] Anthropic just released Phase 2 of Project Fetch. They gave their latest AI model a robotic dog and told it to figure ou... — reactive:ai-coding-agents-robot-training (2026-06-18)
  32. [32] AnthropicAI just released Phase 2 of Project Fetch. They gave their latest AI model a robotic dog and told it to figure ... — reactive:ai-coding-agents-robot-training (2026-06-18)
  33. [33] Nvidia Built Robots That Train Themselves Using AI Coding Agents — reactive:ai-coding-agents-robot-training
  34. [34] Claude just entered the robodog lab. — reactive:ai-coding-agents-robot-training (2026-06-23)
  35. [35] AI agents take over robot training at Nvidia - Techzine Global — reactive:ai-coding-agents-robot-training