AgenticRoboRL lifts 9B robot planner from 18.1% to 71.0%

A new post-training framework called AgenticRoboRL raises a 9B-parameter robot planner's success rate on the LIBERO-PRO benchmark from 18.1% to 71.0% without external memory, approaching the 72.1% that Codex (GPT-5.5) achieves on the same benchmark with external memory, according to a paper posted to arXiv on October 20.

The authors' framing is that robotics has been pulling the wrong lever. "Recent work improves these systems through better harnesses and external memory, but relies on powerful closed-source models as planners and does not improve the planners' own capabilities," they write. Coding agents, by contrast, have gained most of their recent capability by post-training the model itself. The paper applies that recipe to robot planners.

The training recipe combines "supervised fine-tuning with multi-turn reinforcement learning." The twist is harness-specific adapters: because robot planners drive execution in different ways, including VLA skill invocation and control-code generation, the adapters connect recorded planner interactions with task outcomes so a common RL pipeline can train across those harnesses while leaving the execution components fixed.

The team says it "will release the framework and trained planner checkpoints." The abstract reports results only on LIBERO-PRO, and gives no real-robot evaluation, no compute budget, and no identification of the 9B base model.