English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

NVIDIA ENPIRE: Eight Codex Agents Run Autonomous Robotics Research in the Physical World

Forum topic · QianXun · 2026-06-18

Summary

On June 17, 2026, NVIDIA GEAR Lab unveiled ENPIRE, described by Jim Fan as the first realization of Physical AutoResearch: autonomous research carried out by AI agents on real robots. The system assigns eight OpenAI Codex agents to eight robots, each with fixed GPU and token budgets, and a single goal of completing tasks while keeping the robots busy, with no human intervention once launched. ENPIRE's key engineering contributions include: (1) hardware-enforced safety via hard kinematic limits and torque-limited compliant grippers, rather than prompt-based instructions; (2) a frozen success classifier trained on short human demonstrations that prevents agents from gaming their own reward; and (3) telemetry tracking robot utilization (MRU), token throughput (MTU), and GPU utilization to compute Tokens-to-Success and Time-to-Success metrics. Benchmarked on cable tying, needle arrangement, and GPU installation tasks, eight-robot parallel exploration significantly outperformed single-robot setups. The system will be open-sourced. Open questions remain around reward-freezing robustness, industrial cost, and comparisons with Pi0, GR00T, and Helix.

What Happened

On June 17, 2026, NVIDIA GEAR Lab released the ENPIRE system. Jim Fan announced it the same day, calling it the first realization of Physical AutoResearch — autonomous research in the physical world. ENPIRE gives 8 Codex agents 8 robots, fixed GPU allocations, and token budgets, with a single objective: complete tasks as fast as possible while keeping the robots busy. Once launched, it runs unattended.

Deep Dive: The Hard Part Is Everything Before Pressing Enter

Jim Fan himself admitted it sounds easy but the engineering is hell. Three components are worth unpacking:

1. Safety Guardrails

For overnight unattended runs, safety can no longer be a polite note in a system prompt. ENPIRE welds safety into two hardware layers:

  • Hard kinematic limits — crossing a boundary triggers task failure and automatic reset.
  • Torque-limited compliant grippers — misaligned insertion causes a stall instead of crushing the robot or objects.
  • The guiding principle: be conservative, so operators can actually sleep.

    2. Defining /done

    If an agent can modify its own reward function, it will inevitably hack the score. ENPIRE's fix is to bolt down the goalposts first: collect a few minutes of success + failure demonstrations, have the agent write a visual classifier to judge outcomes, hill-climb until stable, then freeze the classifier and embed it in the Gym environment — nobody touches it during the research run.

    3. System Telemetry

    Robot-seconds are the scarcest resource, GPU-seconds next, tokens last. All three are instrumented and fed back to ENPIRE for real-time resource awareness. Three metrics:

  • MRU — fraction of time the robot is active
  • MTU — tokens consumed per minute, reflecting whether the agent is genuinely thinking
  • GPU utilization
  • These combine into two budget-to-output ratios: Tokens-to-Success and Time-to-Success.

    Experiments

    Three tasks were selected: cable tying, needle arrangement, and GPU installation. Conclusion: 8-robot parallel exploration is significantly faster than a single robot. The system will be open-sourced.

    Why It Matters

    1. It pushes LLM agents to the scarcest position in the physical world — the robot embodiment — and closes the loop. Agents previously did AutoResearch in sandboxes (math, code); ENPIRE pulls that path onto real hardware. 2. Resource budget awareness. Three utilization metrics (robot, token, GPU) become observable for the first time, meaning the agent itself is now being "accounted for" — ROI discipline like human engineering. 3. Safety is welded, not prompted. Two hardware protection layers make it truly possible to "press Enter and walk away" — the minimum bar for genuine unattended operation.

    Risks and Open Questions

  • How long can the reward-freezing mechanism hold? For slightly more complex tasks, the visual classifier itself may fall short.
  • Does 8-robot parallelism multiply energy and maintenance costs beyond what industrial deployment can sustain?
  • Details on OpenAI Codex's role in the physical agent are undisclosed — is it an LLM controller, or actually modifying an RL policy?
  • Paper/code release timing is unannounced; cross-lab reproduction will be difficult.
  • No comparison yet with Pi0, GR00T, or Helix — the physical intelligence field sits at a fork between "model-first" and "agent-first"路线.
Source: https://x.com/DrJimFan/status/2067283904986517866

Tags

#nvidia#enpire#codex#robotics#autonomous-research#jim-fan#embodied-ai#physical-autoresearch

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/177981466