ROBBRIEF
On the Push T benchmark, where a robot pushes a T shaped block into a target pose, an AI coding agent wrote a policy from scratch. It hit 100% success and used 46% fewer steps than a baseline trained on 200 human demos.
An AI coding agent taught itself a classic robot pushing task without a single human demonstration, according to a new arXiv preprint. The agent located the Push-T benchmark online, a setup where a robot uses a single point of contact to push a T-shaped block into a target pose, then wrote and iterated an algorithmic policy until it solved the task.
The result: 100% success on the task and roughly 46% fewer steps than the strongest baseline, a diffusion policy trained on 200 human demonstrations. The comparison comes from a single configuration of that baseline, on a low-dimensional, single-point-of-contact simulation, so the win is real but narrow.
The same agent extended the trick from the T shape to the full alphabet (Push-A through Push-Z) using a self-generated curriculum, and wrote simulation code for Franka and UR5 robot arms working in 3D, per the paper.
The whole policy was authored as code, the way a developer writes an algorithm. No human demonstrations were collected; the agent only iterated against its own simulation.
Replication has not happened, broader task variety has not been tested, and whether the gains hold outside Push-T is unknown. The result is simulated, measured against a single baseline configuration by the paper's own authors.