CoRL workshop: running Karpathy's autoresearch on real robots — the sandbox, not the model, is the bottleneck

Kyrannio · x · 2026-11-19

From the CoRL 2026 Workshop on Agentic Robotics (the poster co-organized it): when asked what it takes to replicate @karpathy's autoresearch — which hillclimbs nanochat overnight — on real robots, CMU's @GuanyaShi argued the model isn't the bottleneck; the sandbox is.

A simulator is a perfect sandbox: free resets, clean rewards, near-zero cost per trial. Shi once ran autoresearch over a weekend on a pushT sim, using plain code as the policy with no neural network. The real world offers none of that, so he presented two ways to close the loop — one being ENPIRE, which runs autoresearch directly on real robots, starting from 50 demos and an empty codebase where the agent builds a fixed code policy — plus a new evaluation exam. The workshop featured collaborators including Unitree Robotics.

Original post →

More from coding & agent

coding & agent channel →