Precise assembly has remained difficult for multi-fingered robots because the tasks are contact-rich and sparse-reward, making imitation data collection and direct reinforcement learning impractical. The authors of Play2Perfect argue that robots should first learn to play with diverse objects and goals, acquiring general skills such as grasping, in-hand reorientation, and pose reaching, before finetuning on assembly.
The framework uses task-agnostic RL pretraining followed by assembly-specific finetuning, which focuses exploration on the final high-precision interactions. The paper systematically studies design choices including object diversity, training objective, trajectory diversity, and goal precision. Their pretrained prior is 33x more sample-efficient than RL from scratch, even when the latter is given dense, multi-stage rewards.
In experiments, Play2Perfect achieved 60% success on tight insertions with only 0.5 mm contact clearance and over 50% success on long-horizon multi-part assembly and screwing, all with zero-shot sim-to-real transfer. These results suggest that play pretraining could reduce the need for specialized grippers, tool attachments, and environment fixtures in precise manipulation tasks.