Play2Perfect: What Matters in Dexterous Play Pretraining for Precise Assembly?

2026-07-11 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Robotics & Autonomous Systems, Artificial Intelligence & Machine Learning · Depth: Expert, extended

Summary

Play2Perfect is a novel reinforcement learning framework designed to enable multi-fingered robots to perform precise, contact-rich assembly tasks. It addresses the challenges of sparse rewards and difficult data collection by employing a two-stage approach: task-agnostic "play" pretraining followed by specialized finetuning. The pretraining phase teaches reusable manipulation priors like grasping and in-hand reorientation on diverse objects. This pretrained policy is then "perfected" on specific assembly tasks, such as tight insertions with 0.5 mm clearance, multi-part assembly, and screwing. Play2Perfect demonstrates 33x greater sample efficiency compared to training from scratch, achieving high success rates within 2-5 hours of RL training. It also exhibits robust zero-shot sim-to-real transfer, with 60% success on tight insertions and over 50% on multi-part assembly and screwing using a 22-DoF Sharpa hand on a 7-DoF KUKA iiwa 14 arm.

Key takeaway

For Robotics Engineers developing dexterous manipulation systems for complex assembly, you should consider adopting a pretraining-finetuning paradigm like Play2Perfect. This approach significantly reduces the sample efficiency burden of sparse-reward tasks, enabling robust sim-to-real transfer for contact-rich operations. By focusing pretraining on precise in-hand manipulation across diverse objects, you can achieve high success rates on challenging tasks like 0.5 mm clearance insertions, avoiding brittle solutions from scratch training.

Key insights

Dexterous robots achieve precise assembly by pretraining general manipulation skills through "play" before task-specific finetuning.

Principles

• In-hand manipulation is critical for transfer.
• Object diversity improves finetuning stability.
• Precise play goals yield better priors.

Method

Play2Perfect trains a goal-conditioned RL policy in simulation on diverse primitive objects to random 6D poses, then finetunes this prior using sparse rewards derived from CAD designs for precise assembly.

In practice

• Use 6D pose-reaching objectives for pretraining.
• Employ online random goal trajectories for diversity.
• Set play goal precision to 1 cm for tight clearances.

Topics

• Reinforcement Learning
• Dexterous Manipulation
• Robotic Assembly
• Sim-to-Real Transfer
• Pretraining
• In-Hand Reorientation

Best for: Research Scientist, AI Scientist, Robotics Engineer, Machine Learning Engineer

Related on AIssential

• Play2Perfect: What Matters in Dexterous Play Pretraining for Precise Assembly? — Pretraining dexterous robots through diverse "play" significantly improves sample efficiency and success in precise assembly.
• Mana: Dexterous Manipulation of Articulated Tools — Mana reinterprets dexterous articulated tool manipulation as an animation problem, enabling zero-shot sim-to-real transfer via a coarse-to-fine pipeline.
• A Minimalist Retargeting-Guided Reinforcement Learning Recipe for Dexterous Manipulation — REGRIND uses retargeted human demos and residual RL for zero-shot sim-to-real dexterous manipulation.
• Contrastive Action-Image Pre-training for Visuomotor Control — CAIP learns visuomotor representations by contrastively pre-training on human hand poses from large-scale egocentric video.
• ContactExplorer: Contact Coverage-Guided Exploration for General-Purpose Dexterous Manipulation — CCGE leverages state-conditioned contact coverage to enable efficient, general-purpose exploration in dexterous manipulation.
• Beyond Binary: Sim-to-Real Dexterous Manipulation with Physics-Grounded Contact Representation — Center-of-Pressure (CoP) enables robust sim-to-real dexterous manipulation by preserving dense, physics-grounded tactile information.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.