AgenticRoboRL
Preprint · 2026

AgenticRoboRLBringing Coding-Agent Post-Training to Agentic Robotics

Xinxin Liu1Bingxin Xu2Wenrui Bao1Ming Li1Wentao Bao3Chen Chen1Yuzhang Shang1

1University of Central Florida2University of Southern California3Michigan State University

Paper · arXiv 2610.28589 Code · GitHub Model · Hugging Face

AgenticRoboRL is the first general post-training framework for agentic robotics systems.

18.1%+52.971.0%
LIBERO-PRO
original 9B → post-trained 9B, no external memory
19.2%+53.072.2%
LIBERO-Plus
same two checkpoints, no extra training
Overview of AgenticRoboRL: a planner and a robot harness interact with the environment over multiple turns, the planner is post-trained with SFT and multi-turn RL, and on LIBERO-PRO a 9B planner rises from 18.1% to 71.0%.
Overview of AgenticRoboRL. (a) Agentic robotics systems combine planners with robot harnesses for multi-turn interaction with the environment. (b) We post-train planners through SFT on successful trajectories followed by multi-turn RL. (c) On LIBERO-PRO, post-training improves a 9B planner’s success rate from 18.1% to 71.0%, approaching the 72.1% achieved by Codex (GPT-5.5) with memory. Mem denotes external memory.
Method

Training pipeline

AgenticRoboRL training pipeline: successful trajectories become SFT examples, the planner interacts with the environment through a harness while a proxy records its calls, and recorded calls are matched with episode outcomes for RL updates.
AgenticRoboRL training pipeline. (a) Successful trajectories are converted into training examples for SFT. (b) The planner interacts with the environment through a robot harness, while a proxy records its model calls. (c) Recorded calls are matched with episode outcomes for RL updates, excluding episodes invalidated by infrastructure failures. Harness-specific adapters support a common training pipeline across different harnesses. Only the planner is trained; execution components remain fixed.
Harnesses

Harnesses

RPent · VLA skill invocation

The planner calls skills

Perception tools locate objects. A frozen VLA policy executes pick skills, and motion tools move the arm. Evaluated on LIBERO-PRO and LIBERO-Plus.

segmentback_projectpi0_pickmove_torelease
CaP-X · control-code generation

The planner writes code

The planner writes executable programs over perception and control APIs, and a motion controller runs them. Evaluated on the Object, Goal, and Spatial suites of LIBERO-PRO.

pythonperception APIcontrol API

The training pipeline does not depend on how a harness executes the planner. Connecting a new harness only needs a new adapter. We show two harnesses that execute plans in different ways.

Episodes

Base vs. SFT vs. RL

Two LIBERO-PRO episodes, position swap

Evaluation recordings, 256 × 256. Tool counts are taken from the planner's recorded calls. Single episodes are examples. Outcomes of individual episodes vary between repeated evaluations, so the 800-episode results below are the measure.

LIBERO-PRO by suite

T is task perturbation, S is position swap. 100 episodes per suite.

Original 9B+ SFT+ SFT + RLsuccess rate, %

LIBERO-Plus, no extra training

10,030 episodes over seven perturbation types.

Original+ SFT+ SFT + RL
Cite

BibTeX

@article{liu2026agenticroborl,
  title   = {AgenticRoboRL: Bringing Coding-Agent
             Post-Training to Agentic Robotics},
  author  = {Liu, Xinxin and Xu, Bingxin and Bao, Wenrui and
             Li, Ming and Bao, Wentao and Chen, Chen and
             Shang, Yuzhang},
  journal = {arXiv preprint arXiv:2610.28589},
  year    = {2026}
}