NVIDIA ENPIRE: eight Codex agents run a robotic laboratory
NVIDIA unveiled ENPIRE: eight Codex AI agents were given robots, GPUs, and the task of improving real actions in a physical laboratory
The NVIDIA, CMU, and UC Berkeley team released both the project and a 29-page paper on a system for robotic AutoResearch. The agents write code, run robots, inspect logs, change the control policy, check the outcome, and try again. The experiments cover tasks like inserting pins into holes, handling zip ties, and plugging a GPU into a slot.
"We give eight Codex agents a fleet of robots, GPU allocation, and a generous token budget, meaning a large amount of language-model labor. The goal is simple: solve the task as quickly as possible, keep the robots busy but safe, and avoid wasting compute," Jim Fan of NVIDIA wrote.
On 16 June, the ENPIRE page described this more calmly than the X post did: a robotics task is turned into an environment that a coding agent can interact with almost like a software test. The agent resets the scene, launches an attempt, receives a success check, opens video and logs, changes the code, and launches another attempt.
To do that, the team first builds the Environment: automatic reset, success checks, safety boundaries, and logging. Then comes Policy Improvement: the agent changes the control-policy code and tries heuristics, imitation learning, or reward-based learning. Rollout runs real attempts on one or several robots. Evolution lets multiple agents compare branches, borrow successful ideas from one another, and discard failed ones through Git.
The paper describes eight identical stations with YAM bimanual robots: each has two arms, cameras, and a dedicated workstation with an RTX 5090. The agent gets a sandbox with a repository, data from the current run, internet access, and permission to execute commands without asking for approval at every step. The result is a small compute cluster where coding tasks are joined by motors, cameras, friction, failures, and costly physical attempts.
The authors report several measurable results. NVIDIA says on the site that ENPIRE policies reached 99% pass@8 on the showcased tasks. In the paper, the team writes that for pin insertion, scaling from one agent to eight cut the time to near-100% success from more than 1.5 hours to about 40 minutes. On Push-T, time to a normalized score of 1.0 fell from about five hours to two.
The physical world immediately spoils the neat picture. In simulation, all tested agents solved Push-T, but on the real robot, two out of three failed. The robots and GPUs still remain underutilized: the agent reads logs, writes code, waits for the model, argues with other branches, and the hardware sits idle. As the setup scales to eight agents, token consumption grows faster than linearly. More robots buy speed, but they buy it with expensive token use and coordination overhead.
At Karpaty, AutoResearch changed a training file, ran a test, and saved the edit based on a numerical metric. At Gladstone, the thinking microscope chose an intervention on living neurons, measured the response, and selected the next experiment. ENPIRE places a robot farm between those two examples: the agent starts spending real physical time. That kind of agent already has a bench, cameras, manipulators, force limits, an emergency reset, a queue of hypotheses, and a cost for every attempt.
In biology, elegant hypotheses run into the wet lab. The team has to run the experiment, return the system to its initial state, understand the result, preserve the setup, and decide whether the robot's next night is worth spending. If systems like this become reliable, power will shift to those who set the goal, write the constraints, accept the validation, and pay for physical mistakes.