Evolutionary search on Graph-as-Policy robot controllers yields 5.27x throughput for $37

matei_zaharia · x · 2026-09-28

Following his Robotics Harness Optimization (RHO) paper with AMD—using evolutionary search to improve robot policies that are whole code repositories—Karim applied RHO to NVIDIA's Graph-as-Policy (GaP, arXiv:2607.05369), where humans write skill nodes (observe, perceive, grasp, move, release) and an LLM wires them into a directed computation graph. After 50 generations of evolutionary search on an example graph: 5.27x throughput gain over baseline, grasps per trial down from 4.17 to 1, and total search cost of just $37 across both experiments. GaP's graph policies are readable, diffable source code with locally editable, searchable stages, letting search propose "add a node" moves without demonstrations. Matei Zaharia amplified the work.

Original post →

More from Embodied

Embodied channel →