AgenticRoboRL post-trains robot planners, lifts 9B model to 71.0% on LIBERO-PRO
An arXiv paper submitted on 20 October 2026 introduces AgenticRoboRL, a framework for post-training robot planners across different robot harnesses by combining supervised fine-tuning with multi-turn reinforcement learning. Harness-specific adapters connect recorded planner interactions with task outcomes so a common RL pipeline can be used while execution components stay fixed. The authors report that on LIBERO-PRO, AgenticRoboRL raised a 9B-parameter planner's success rate from 18.1% to 71.0% without external memory, approaching the 72.1% they attribute to Codex (GPT-5.5) with external memory; these figures are self-reported and not independently verified. The authors say they will release the framework and trained planner checkpoints.
The paper argues that post-training the planner itself, rather than relying on powerful closed-source models plus better harnesses and external memory, is an effective way to improve robot planners.
Computer Science > Robotics
arXiv:2610.28589 (cs)
[Submitted on 20 Oct 2026]
Title:AgenticRoboRL: Bringing Coding-Agent Post-Training to Agentic Robotics
Authors:Xinxin Liu, Bingxin Xu, Wenrui Bao, Ming Li, Wentao Bao, Chen Chen, Yuzhang Shang
View a PDF of the paper titled AgenticRoboRL: Bringing Coding-Agent Post-Training to Agentic Robotics, by Xinxin Liu and 6 other authors