Glonce

AgenticRoboRL post-trains a 9B robot planner to 71.0% on LIBERO-PRO

researchThe Glonce editorsHugging Face🔥58Glonce rating58Impact45Novelty60Relevance95Surprise30Players10Credibility70
AgenticRoboRL post-trains a 9B robot planner to 71.0% on LIBERO-PRO

A new paper introduces AgenticRoboRL, a framework for post-training robot planners across different robot harnesses. Agentic robotics systems pair a planner with a robot harness to carry out complex manipulation tasks. The authors say recent work improves these systems through better harnesses and external memory, but still relies on powerful closed-source models as planners and does not improve the planners' own capabilities.

Their starting point is coding agents, which have improved substantially by post-training the model itself. The paper asks whether the same move can improve robot planners. The obstacle is that harnesses differ in how they execute plans. The authors name two mechanisms: VLA skill invocation and control-code generation.

AgenticRoboRL's training recipe combines supervised fine-tuning with multi-turn reinforcement learning. To cope with the different execution mechanisms, harness-specific adapters connect recorded planner interactions with task outcomes. That lets all harnesses share one common RL training pipeline while the execution components stay fixed.

The headline result is on LIBERO-PRO. AgenticRoboRL raises the success rate of a 9B-parameter planner from 18.1% to 71.0% without external memory. For comparison, Codex (GPT-5.5) with external memory achieves 72.1%, so the trained planner approaches that score but does not exceed it. The authors conclude that post-training is an effective approach for improving robot planners, and say they will release the framework and trained planner checkpoints.

Key facts

  • AgenticRoboRL is a framework for post-training robot planners across different robot harnesses, combining supervised fine-tuning with multi-turn reinforcement learning.
  • On LIBERO-PRO, a 9B-parameter planner goes from 18.1% to 71.0% success without external memory.
  • That approaches, but does not exceed, the 72.1% reported for Codex (GPT-5.5) with external memory.
  • Harness-specific adapters link recorded planner interactions to task outcomes, so one RL pipeline works while execution components stay fixed.
  • The authors say they will release the framework and trained planner checkpoints.

Why it matters

Prior agentic robotics work, as the authors describe it, improves harnesses and external memory while leaning on powerful closed-source models as planners, leaving the planner itself unchanged. This paper borrows the approach that has worked for coding agents, post-training the model itself, and applies it to robot planners. The reported jump on LIBERO-PRO, from 18.1% to 71.0% for a 9B planner, is the evidence offered that the approach works.

Who it affects

Robotics researchers building planner-plus-harness systems are the main audience, especially those whose harnesses differ in execution style, such as VLA skill invocation or control-code generation. Teams that depend on closed-source models as planners are the ones the authors position the work against.

How to use it

Nothing is usable yet. The authors say they will release the framework and trained planner checkpoints, but the abstract gives no release date, license or repository link. The design to note is the adapter approach: a harness-specific adapter connects recorded planner interactions with task outcomes, so a common RL pipeline can train planners while the execution components stay fixed.

How solid is it

This is a preprint, and the account here rests on the abstract alone. The central numbers are specific: 18.1% to 71.0% on LIBERO-PRO for a 9B planner without external memory, against 72.1% for Codex (GPT-5.5) with external memory. The conclusion that post-training is effective is the authors' own. The abstract gives no training compute, data size or reward design details, and the code and checkpoints are promised rather than released.

Risks and caveats

The comparison is not like for like: the trained planner runs without external memory, while the 72.1% Codex figure uses it. The result approaches that score rather than beating it. The abstract reports numbers only for LIBERO-PRO; it gives none for other benchmarks, real robots or other harnesses. The 9B planner's base model is not named, so the source does not say what was post-trained.

“These results demonstrate post-training as an effective approach for improving robot planners.”

— AgenticRoboRL paper abstract