Playful Agentic Robot Learning

The article introduces Playful Agentic Robot Learning, focusing on Robotics Agent Teams (RATs) that utilize self-directed play for skill acquisition prior to task execution. The study demonstrates that skills learned during play significantly enhance performance on downstream tasks, achieving improvements of 20.6 and 17.0 percentage points over baseline models in LIBERO-PRO and MolmoSpaces, respectively. This approach allows for the integration of learned skills into existing Code-as-Policy frameworks, enhancing performance in RoboSuite and real-world applications by 8.9 and 8.8 points without requiring model fine-tuning, which is crucial for practitioners aiming to develop more adaptable and efficient robotic systems.