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

AgenticRoboRL. Bringing Coding-Agent Post-Training to Agentic Robotics. By Xinxin Liu and six others, from the University of Central Florida, the University of Southern California, and one other institution. Published on the twentieth of October. Its code repository is worked out: about forty-five source files, with a readme and a licence.

Agentic robotics systems pair a planner with a robot harness to carry out complex manipulation tasks. Recent work has improved these systems by building better harnesses and adding external memory. But it relies on powerful closed-source models as the planners, and it leaves the planners' own capabilities where they were. Everything around the planner gets better while the planner itself is treated as fixed, a component bought rather than built. Coding agents went a different way. They improved substantially by post-training the model itself, so that the gains live in the weights rather than in the tooling around them.

AgenticRoboRL asks whether the coding-agent route works for robots too, by post-training the robot planner directly. The obstacle is that harnesses differ in how plans get executed. Some invoke skills from a vision-language-action model. Others generate control code. A training framework has to accommodate both.

The recipe combines supervised fine-tuning with multi-turn reinforcement learning. To handle the different execution mechanisms, harness-specific adapters connect recorded planner interactions to task outcomes. That gives every harness a common RL training pipeline while the execution components themselves stay fixed. Only the planner learns. Whether a harness calls a skill or runs generated code, its adapter turns what happened into the same kind of training signal: a record of what the planner did, joined to whether the task succeeded. From there the reinforcement learning is shared.

On LIBERO-PRO, AgenticRoboRL raises the success rate of a nine billion parameter planner from eighteen point one percent to seventy-one percent, with no external memory. For comparison, Codex running GPT-5.5 reaches seventy-two point one percent, and it has external memory to draw on. The authors say they will release the framework and the trained planner checkpoints.

The result reads as a transfer of strategy from one field to another. Where robotics had been improving the scaffolding around a fixed closed model, this work improves the model inside the scaffolding, and a small open planner trained that way comes within about a point of a frontier system that also has memory.