Import AI 463: Self-improving robots; a 10k Chinese GPU cluster; and an elegiac essay for the human era

NVIDIA researchers have put together something that looks a lot like the software scaffolding behind autonomous coding agents, except instead of writing code it's controlling robot arms. The system, called ENPIRE, gives physical robots the same kind of loop that agentic coders already enjoy: try something, check if it worked, reset the scene, and try again, with almost no human in the middle. It's an early, clunky sketch of what it might look like for AI to start improving itself out in the real world rather than just inside a training cluster.

The architecture has four pieces. An Environment module resets the physical setup after every trial. A Policy Improvement module decides how to refine the robot's behavior. A Rollout module runs the policy on real hardware, sometimes across several robots at once. And an Evolution module has the coding agent read logs, dig through research literature, and rewrite its own training code to fix whatever's going wrong. Put together, this turns messy real-world robot learning into something closer to a controllable optimization job, which is the whole point: minimize human babysitting, maximize fair comparisons between different training recipes.

The hardware is refreshingly modest. Each station pairs two YAM manipulator arms from I2RT with a scattering of cameras and a single workstation running an RTX 5090. Nothing exotic. What's more interesting is what NVIDIA got out of it: frontier coding agents autonomously developed policies that hit a 99% success rate on tricky tasks like the PushT benchmark, sorting pins into a box, and slicing a zip tie with a cutter. They even tried having a robot insert GPUs into a motherboard, which feels almost too on-the-nose given who funded the research.

Not every model performed the same. GPT-5.5 running inside Codex traded blows with Opus 4.7 inside Claude Code for top spot, while Kimi-2.6 fell behind. Throwing more agents at a problem helped too — eight agents working in parallel reached better solutions faster than a lone agent, and sometimes beat it outright, presumably because more agents means more of the solution space gets explored before anyone commits.

But scaling up isn't free. NVIDIA found that as more robots joined the fleet, the agents spent more time reading logs, writing code, and waiting on the language model, which meant robot utilization actually dropped even as GPU usage climbed. Adding robots to this kind of loop doesn't parallelize as cleanly as adding GPUs to a training run, and that mismatch is likely to be the real bottleneck long before anyone worries about the robots getting too clever for their own good.