AgenticRoboRL: Bringing Coding-Agent Post-Training to Agentic Robotics
To this end, we introduce AgenticRoboRL, a framework for post-training robot planners across different harnesses.
ProofPaper ↗
Key points
- Agentic robotics systems combine planners with robot harnesses to perform complex manipulation tasks.
- We therefore investigate whether post-training can also improve robot planners.
- The training recipe combines supervised fine-tuning with multi-turn reinforcement learning.
- On LIBERO-PRO, AgenticRoboRL raises the success rate of a 9B-parameter planner from 18.1% to 71.0% without external memory, approaching the 72.1% achieved by Codex (GPT-5.5) with external memory.
Sources (2)
- [1]AgenticRoboRL: Bringing Coding-Agent Post-Training to Agentic RoboticsHugging Face Daily Papers · Oct 20, 12:00 AM
To this end, we introduce AgenticRoboRL, a framework for post-training robot planners across different harnesses.
Agentic robotics systems combine planners with robot harnesses to perform complex manipulation tasks.
- [2]AgenticRoboRL: Bringing Coding-Agent Post-Training to Agentic RoboticsarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 20, 05:41 PM · same content
Extractive summary: sentences quoted from the sources.
Before this
- Oct 19, 2026HAF-VLA: Hierarchical Action Flow Framework for Humanoid Vision-Language-Action Control
- Oct 19, 2026BetaMem: Learning to Remember a Changing World for Embodied Agents
- Oct 19, 2026SPAR: Structural Pruning and Action Recovery for Vision-Language-Action Models
- Oct 19, 2026ClueFlow: Learning Long-Horizon Search through Attributed Research-State Transitions
- Oct 19, 2026HarnessRL: Co-Evolving Policies with a Just-in-Time Training Harness for Open-Ended RL
- Oct 19, 2026HarnessPro: Recursive Harness Self-Improvement with Pareto Optimization for Performance-Cost Trade-offs