Playful Agentic Robot Learning

2026-06-17

AI Summary

Existing embodied agents rely on task instructions to acquire skills, lacking the capacity for autonomous, continuous learning. This work proposes an embodied code-agent framework that introduces, for the first time, a self-driven play mechanism enabling agents to proactively explore, execute, and distill reusable code-based skills during instruction-free phases, thereby achieving pre-acquired skill accumulation and plug-and-play transferability. Built upon the Robot Agent Teams (RATs) architecture, the framework integrates task generation, code-based policy planning, feedback-driven retry, and skill distillation. Experiments demonstrate that the approach outperforms baselines by 20.6 and 17.0 percentage points on LIBERO-PRO and MolmoSpaces, respectively. Moreover, the resulting skill library significantly enhances the performance of other Code-as-Policy agents in both RoboSuite simulations and real-world tasks, yielding improvements of +8.9 and +8.8 points.

Abstract

Current agentic robot systems can write executable Code-as-Policy programs, observe feedback, and revise behavior across multiple attempts, but they remain largely task-driven: reusable skills are acquired only after explicit instructions. We study Playful Agentic Robot Learning, where an embodied coding agent uses self-directed play as a continual skill-learning stage before downstream tasks arrive. We introduce RATs, Robotics Agent Teams designed for play-time skill acquisition. During play, RATs proposes novel yet learnable exploratory tasks, plans and executes robot-code policies, verifies intermediate progress, diagnoses failures, retries with dense, step-level feedback, and distills successful executions into a persistent code skill library. At test time, the agent reuses relevant skills from this frozen library to help solve new tasks. Experiments in LIBERO-PRO and MolmoSpaces show that play-learned skills improve held-out downstream tasks over no-play and random-play baselines, with 20.6 and 17.0 percentage-point gains over CaP-Agent0 on LIBERO-PRO and MolmoSpaces, respectively. Moreover, the learned skills can be plugged into other inference-time Code-as-Policy agents by simply retrieving them into the context, improving RoboSuite and real-world transfer by 8.9 and 8.8 points, respectively, without finetuning the underlying model.

Problem

• Playful Learning
• Agentic Robot
• Skill Acquisition
• Code-as-Policy
• Embodied AI

Innovation

• Playful Agentic Robot Learning
• Code-as-Policy
• Skill Library
• Self-directed Play
• Robotics Agent Teams