GaP paper: graph-as-policy multi-agent harness makes robot automation reliable

rbhar90 · x · 2026-09-08

A Berkeley/Nvidia CoRL paper introduces Graph-as-Policy (GaP), a multi-agent coding harness that builds directed computation graphs (perception, planning, control) from the MORSL skill library and rehearses task graphs in internal simulation to close the reliability gap of model-free policies on Variational Automation tasks. Discussion: Goldberg notes harness design like GaP can further boost benchmarked models such as Astra; Letian Fu argues harnesses + tool calls (IK, SAM3) beat raw VLA on cost, speed and reliability, and that self-evolving harnesses and skill libraries will rapidly expand the capability Pareto frontier — robotics as the next frontier of multimodal agents.

Original post →

More from Embodied

Embodied channel →