AgenticRoboRL: Bringing Coding-Agent Post-Training to Agentic Robotics

Key points

It combined supervised fine-tuning with multi-turn reinforcement learning to train only the robot planner while keeping the VLA execution components fixed.
Harness-specific adapters connect recorded planner calls with episode outcomes, so a single training pipeline supports both VLA skill invocation and control-code generation.
On LIBERO-PRO, SFT raised the 9B planner's success rate from 18.1% to 65.1%, and RL further raised it to 71.0%.
On the 10,030 LIBERO-Plus evaluations unseen in training, the same checkpoint improved from 19.2% to 72.2%.
With the CaP-X harness, it improved the success rate from 2.7% to 28.5%, exceeding the 27B teacher's 20.3%.
The RL reward used only task success, with no intermediate rewards or a length penalty.

External research summaries. These are not HDATF publications or measured product results.