Event
AgenticRoboRL: Cross-Embodiment Post-Training of Robot Planners, Success Rate Rises to 71%
1 reports · 1 sources
The AgenticRoboRL paper author team proposes a framework for post-training planners across different robot embodiments, reporting that on LIBERO-PRO it raises the success rate of a 9B-parameter planner from 18.1% to 71.0% (without external memory), approaching Codex (GPT-5.5) with external memory at 72.1%. The training recipe combines supervised fine-tuning with multi-turn reinforcement learning; an institution-specific adapter connects the planner's interaction records and task outcomes to a unified RL training pipeline while keeping the execution components fixed. The authors state that the framework and trained planner checkpoints will be released.
Report timeline
- AgenticRoboRL brings coding-agent post-training to robot planners
AgenticRoboRL proposes a framework for post-training robot planners across different robot harnesses. Its training recipe combines supervised fine-tuning with multi-turn reinforcement learning, and harness-specific adapters connect planner interaction records and task outcomes to a unified RL training pipeline.
arXiv cs.RO ·