Playful Agentic Robot Learning

## The Core Innovation

Robots today learn tasks reactively—they need explicit instructions to acquire skills. This paper flips that model: robots should *play first*, building up a reusable skill library through self-directed exploration, then tap that library when real tasks arrive. The result is substantial: 20.6 percentage-point improvement over the baseline agent on LIBERO-PRO benchmarks, and the learned skills transfer to other agents without retraining.

## How Playful Learning Works

The system, called RATs (Robotics Agent Teams), operates in two phases.

**Play phase:** The agent doesn't wait for task instructions. Instead, it proposes exploratory tasks to itself—things like "move the gripper to the table corner" or "push the object forward." For each self-generated task, the agent writes robot-control code (Code-as-Policy, meaning the policy is executable program text), executes it on the robot, observes what happened, and decides whether it succeeded. Crucially, when attempts fail, the system diagnoses *why*—parsing error messages, visual feedback—and retries with fine-grained hints. Successful executions get saved as reusable code snippets into a persistent skill library.

**Test phase:** When a downstream task arrives (e.g., "arrange objects on a shelf"), the agent retrieves relevant skills from its play-learned library, incorporates them into the context, and uses them to accelerate solving the new task. Importantly, the underlying language model stays frozen; skills are just added to the prompt.

The key insight is that play generates diverse, low-stakes experience. The agent discovers what works and what doesn't before facing real objectives, building a foundation of verified behaviors.

## Why This Matters

**Practical transfer:** The learned skills aren't locked to one model or one task. Testing shows they improve performance on held-out tasks (+20.6 points on LIBERO-PRO), and they also boost other Code-as-Policy agents by 8.9 points on RoboSuite and 8.8 points on real-world transfer—without any finetuning. This modularity is rare in robot learning.

**Efficiency at scale:** Robots accumulate skills during idle periods (play time) rather than requiring explicit task definitions upfront. As the skill library grows, downstream task solving becomes faster and more reliable.

**Biological plausibility:** The approach mirrors how animals and children learn—play is a natural, low-pressure way to explore and acquire competence before facing real-world demands.

The experiments span simulated benchmarks (LIBERO-PRO, MolmoSpaces, RoboSuite) and real robot setups, consistently showing that play-learned skills outperform both no-play and random-play baselines. The 17–20 percentage-point gains suggest that the self-directed exploration strategy is substantially more effective than passive or undirected learning.