arXiv 2610.28589
AgenticRoboRL: Bringing Coding-Agent Post-Training to Agentic Robotics
Robotics20 upvotes
Authors: Xinxin Liu, Bingxin Xu, Wenrui Bao, Ming Li, Wentao Bao, Chen Chen, Yuzhang Shang
Overview
The authors introduce AgenticRoboRL, a framework for post-training robot planners across different robot harnesses. It combines supervised fine-tuning with multi-turn reinforcement learning, using harness-specific adapters so one RL pipeline can work with both VLA skill invocation and control-code generation while execution components stay fixed.
Introduction
Agentic robot systems rely on planners that are often powerful closed-source models, and prior improvements have focused on harnesses and external memory rather than the planner's own capabilities.
Results
On LIBERO-PRO, AgenticRoboRL raises a 9B-parameter planner's success rate from 18.1% to 71.0% without external memory. This approaches the 72.1% of Codex (GPT-5.5) with external memory.
Conclusion
Post-training the planner itself is an effective way to improve robot agents, reducing dependence on large closed-source models and expensive memory systems.