Researchers have built a robot learning system that improves task performance by letting agents play first and work second.
The paper introduces RATs — Robotics Agent Teams — an approach where embodied coding agents spend a pre-task phase inventing and attempting their own exploratory challenges. During this play period, the system proposes tasks it thinks are novel but achievable, executes robot-code policies, checks its own progress, diagnoses failures, and distills what worked into a persistent skill library. When real tasks arrive, the agent pulls relevant skills from that frozen library rather than starting from scratch. On the LIBERO-PRO and MolmoSpaces benchmarks, play-trained agents outperformed no-play baselines by 20.6 and 17.0 percentage points, respectively. Crucially, the harvested skills transferred to entirely different agents — improving performance on RoboSuite and real-world tests by roughly 8.9 and 8.8 points — without any additional model fine-tuning.
The deeper implication here is architectural: the researchers are arguing that task-driven robot learning leaves capability on the table by skipping an unstructured exploration phase. That mirrors how motor skills develop in biological systems, and it suggests the bottleneck in current agentic robotics may be less about model size and more about how agents spend their compute before deployment.
The obvious caveat is that play time costs something — compute and clock cycles spent before a single real task is attempted. Whether that upfront investment pays off outside controlled benchmarks, in environments that aren't already well-structured for code-as-policy approaches, remains an open question.