arXiv Robotics — research abstracts· Haichao Liu, Yuanjiang Xue, Yuheng Zhou, Li Yang, Haoyuan Deng, Yinan Liang, Lihua Xie, Ziwei Wang·2026-10-27 04:00 UTC· 15 hours agoSignalEditorial score85

UniManip: Open-World Robotic Manipulation via Agentic Operational Graphs

Summary

UniManip is a unified framework for robotic manipulation that uses an Agentic Operational Graph (AOG) to orchestrate tasks, track states, and recover from failures. It achieves state-of-the-art performance on LIBERO-Pro and demonstrates strong zero-shot generalization across various tasks and embodiments, outperforming leading VLA models and hierarchical baselines.

Editorial context

UniManip introduces a novel framework for robotic manipulation by integrating an Agentic Operational Graph (AOG) to bridge high-level semantic intent with low-level physical interaction. The approach enables robust zero-shot generalization across diverse tasks and embodiments, outperforming existing VLA models and hierarchical planners.

Source: arXiv Robotics — research abstracts · Read original article ↗

Loading article text…

Source：arXiv Robotics — research abstracts · arxiv.org

#Research #Robot policies #Imitation learning #Reinforcement learning #World models #Sim-to-real