Playful Agentic Robot Learning

Researchers introduce RATs (Robotics Agent Teams), an agentic robot learning system that uses self-directed play to acquire reusable skills before receiving downstream tasks. The approach demonstrates significant performance improvements on robotics benchmarks and enables learned skills to transfer across different agents without finetuning.

The research addresses a fundamental limitation in current agentic robot systems: their reliance on explicit task instructions to develop reusable capabilities. Traditional Code-as-Policy approaches require direct supervision, whereas RATs enables autonomous skill discovery through play-based exploration. This represents a meaningful shift toward more self-directed learning paradigms that could reduce engineering overhead in robotics deployment.

The methodology leverages several key innovations: agents propose novel exploratory tasks during play phases, execute robot-code policies, diagnose failures with dense feedback, and systematically distill successful behaviors into persistent skill libraries. This approach mirrors how biological systems learn through unstructured play before specializing for specific tasks. The frozen skill library enables downstream transfer, allowing new agents to retrieve and apply learned behaviors through in-context retrieval without requiring model retraining.

Empirical results demonstrate substantial gains across multiple environments. Performance improvements of 20.6 and 17.0 percentage points over baseline methods on LIBERO-PRO and MolmoSpaces respectively indicate the approach's effectiveness. Notably, the learned skills transfer to entirely different agents and real-world scenarios, with improvements of 8.8-8.9 points without finetuning. This modular capability has practical implications for robotics development, potentially accelerating deployment timelines by enabling knowledge reuse across platforms.

The work's significance extends beyond benchmark performance. By decoupling skill acquisition from task specification, the research opens pathways for more efficient robot learning in resource-constrained environments. Future development should focus on whether this approach scales to longer-horizon tasks and whether emergent skill combinations enable novel downstream applications.

• →RATs uses autonomous play-based exploration to acquire reusable robot skills before receiving explicit task instructions, improving downstream performance by 17-20 percentage points.
• →Learned skills transfer across different agentic agents and real-world scenarios through context retrieval without requiring model finetuning.
• →The approach demonstrates a paradigm shift from task-driven learning toward self-directed skill discovery in embodied AI systems.
• →Performance gains extend to RoboSuite and real-world transfer tasks, suggesting practical applicability beyond simulation environments.
• →The modular skill library architecture enables knowledge reuse across platforms, potentially reducing engineering costs in robotics deployment.