RATs framework lets robots learn reusable skills through self-directed play
Researchers introduce RATs (Robotics Agent Teams), a framework where embodied coding agents acquire reusable skills through self-directed play before any downstream tasks are assigned, yielding significant performance gains over no-play baselines.
Score breakdown
The RATs framework demonstrates that self-directed play — without any explicit task instructions — can build a reusable, transferable code skill library that improves both in-distribution and real-world robot task performance without retraining the underlying model.
- 01RATs stands for Robotics Agent Teams, designed for play-time skill acquisition before downstream tasks arrive.
- 02During play, RATs proposes exploratory tasks, executes robot-code policies, diagnoses failures, and retries with dense step-level feedback.
- 03Successful play executions are distilled into a persistent, frozen code skill library reused at test time.
Junyi Zhang, Jiaxin Ge, and Hanjun Yoo identify a key limitation in current agentic robot systems: although they can write Code-as-Policy programs, observe feedback, and revise behavior across attempts, reusable skills are only acquired after explicit instructions. Their paper proposes Playful Agentic Robot Learning as a remedy — a continual skill-learning stage driven entirely by self-directed play that runs before any downstream task is specified.
The core system, RATs (Robotics Agent Teams), operates in two phases.
The core system, RATs (Robotics Agent Teams), operates in two phases. During play, RATs proposes novel yet learnable exploratory tasks, plans and executes robot-code policies, verifies intermediate progress, diagnoses failures, and retries using dense step-level feedback. Successful executions are distilled into a persistent code skill library. At test time, the library is frozen and relevant skills are retrieved into context to help solve new tasks — a plug-in mechanism that requires no finetuning of the underlying model.
Experiments across LIBERO-PRO and MolmoSpaces demonstrate that play-learned skills improve held-out downstream task performance by 20.6 and 17.0 percentage points over the CaP-Agent0 baseline, respectively, compared to no-play and random-play conditions. The transferability of the skill library is further validated by plugging it into other inference-time Code-as-Policy agents, yielding gains of 8.9 points on RoboSuite and 8.8 points on real-world transfer without any model finetuning.
Key facts
- 01RATs stands for Robotics Agent Teams, designed for play-time skill acquisition before downstream tasks arrive.
- 02During play, RATs proposes exploratory tasks, executes robot-code policies, diagnoses failures, and retries with dense step-level feedback.
- 03Successful play executions are distilled into a persistent, frozen code skill library reused at test time.
- 04On LIBERO-PRO, play-learned skills achieve a 20.6 percentage-point gain over the CaP-Agent0 baseline.
- 05On MolmoSpaces, the gain over CaP-Agent0 is 17.0 percentage points.
- 06Skills can be plugged into other Code-as-Policy agents via context retrieval, improving RoboSuite by 8.9 points and real-world transfer by 8.8 points.
- 07No finetuning of the underlying model is required to transfer the skill library to other agents.
Topics
Summary and scoring are generated automatically from the original article. We always link back to the publisher and never republish images or paywalled content. Last processed Jun 19, 2026 · 10:25 UTC. How this works →