Paper Overview
Field: AI / Robotics Authors: Kaiyuan Chen, Shuangyu Xie, Letian Fu, Justin Yu, William Pacini, Sandeep Bajamahal, Hudson Kim, Jaimyn Drake, Daehwa Kim, Haoru Xue, Jonathan Francis, Christian Juette, Peter Schaldenbrand, Muhammet Yunus Seker, Ruwan Wickramarachchi, Uksang Yoo, Guanzhi Wang, Adithyavairavan Murali, Balakumar Sundaralingam, S. Shankar Sastry, Spencer Huang, Yuke Zhu, Linxi 'Jim' Fan, Ken Goldberg Published: 2026-07-06 arXiv: 2607.05369
Abstract
For robots to work reliably in commercial and industrial applications, can recent advances in agentic programming systems be combined with the open-world adaptability of model-free policies? This paper focuses on variation automation (VA) tasks, where object geometries and poses vary more than in fixed automation. Model-free policies often struggle to bridge the reliability gap for VA tasks.
Inspired by task and motion planning (TAMP) and ROS, the authors propose Graph-as-Policy (GaP), a multi-agent coding framework that generates directed computation graphs with perception, planning, and control nodes from a Modular Open Robot Skill Library (MORSL). GaP constructs internal simulated environments and rehearses task instances across different graphs in parallel, iteratively refining graph structure and parameters to improve success rates and throughput.
GaP was evaluated on eight new open VA benchmark tasks (four in simulation plus four real-world tasks) and significantly outperformed baselines.
Links
- arXiv: 2607.05369
- Project website: https://graph-robots.github.io/gap