[Paper Notes] Play2Perfect: What Matters in Dexterous Play Pretraining for Precise Assembly?
Published:
Play2Perfect takes a simple position on dexterous assembly: an agent should acquire broad manipulation competence before it is asked to solve a millimetre-scale contact problem. Its two-stage recipe first trains one goal-conditioned policy to play with many objects in free space, then finetunes that policy with sparse rewards on a particular CAD-defined assembly task. The pretraining phase learns reusable grasping, in-hand reorientation, and 6D pose-control behaviors; finetuning concentrates exploration on contact, alignment, insertion, and screwing.
The important result is not merely that pretraining helps. The paper identifies what makes a play prior transfer: diverse objects, random goal trajectories, a 6D objective with orientation control, and precise goal tolerances. Each choice pushes the hand toward finger-driven in-hand control, the capability needed when an assembly trajectory can no longer be achieved by moving the arm with a fixed grasp.
Paper Info
“Play2Perfect: What Matters in Dexterous Play Pretraining for Precise Assembly?” is by Tyler Ga Wei Lum, Kushal Kedia, C. Karen Liu, and Jeannette Bohg (Stanford University and Cornell University). It is an arXiv preprint, arXiv:2606.26428, revised in July 2026. The project page provides videos and additional implementation material.
Why Sparse-Reward Assembly Is So Difficult
A dexterous hand starting from a random policy must first discover how to grasp an object, retain it, reorient it, bring it to the fixture, align it, and cope with contact before it reaches an assembly terminal reward. This makes direct sparse-reward reinforcement learning extremely unlikely to obtain useful early trajectories. Dense, task-specific shaping can guide exploration, yet it encodes detailed knowledge about the particular task and may create shortcuts that do not survive perturbations.
Play2Perfect shifts the learning burden. Free-space play makes object motion and grasp recovery attainable, so it can train a reusable prior without knowing the final assembly geometry. Downstream RL then adapts an already dexterous policy to the final high-precision interaction. This keeps the assembly task sparse-reward while avoiding discovery from a completely unskilled initialization.
Stage 1: Goal-Conditioned Dexterous Play
The play policy controls both a 7-DoF KUKA iiwa 14 arm and a 22-DoF Sharpa five-fingered hand. At time (t), it conditions on robot proprioception (s_t), current object pose (o_t\in SE(3)), target pose (g_t\in SE(3)), and bounding-box dimensions (\phi):
[ \pi_\theta(s_t,o_t,g_t,\phi). ]
Simulation procedurally generates cuboids and cylinders whose dimensions fit the hand. Density, extra end masses, center of mass, and inertia are randomized as well. The default pretraining set contains 1,000 objects. This deliberately modest object family makes simulation fast and stable while exposing the policy to geometry and dynamics that cannot be handled with one memorized grasp.
Every episode follows a sequence of random 6D object-pose goals. The first goal requires grasping and lifting from the table. Later goals require maintaining the grasp while changing translation and rotation. The reward combines action smoothness, a lifting incentive, and a goal term with a large success bonus:
[ r = r_{\mathrm{smooth}} + r_{\mathrm{grasp}} + r_{\mathrm{goal}}, \qquad r_{\mathrm{goal}}\text{ is successful when }d_{\mathrm{pose}}(o_t,g_t)<\epsilon. ]
The default (\epsilon=1\,\mathrm{cm}) is consequential. It turns play into accurate pose control instead of coarse transport. Goals are sampled online: the first is broad in workspace and later goals sit near the prior pose with substantial rotations. These trajectories repeatedly force regrasping and finger-mediated orientation changes.
Stage 2: Turn a CAD Assembly into Sparse RL
For a target assembly, the method starts from CAD meshes and the completed part configuration. It uses assembly-by-disassembly: find a feasible sequence of removals, reverse it, and obtain an assembly sequence. Each step inserts a part (p_i) into the fixture formed by already assembled parts (f_i).
The CAD model supplies the desired relative transform (T^{f_i}_{p_i}), so the final goal follows from the current fixture pose:
[ g_i^M = f_i^t T^{f_i}_{p_i}. ]
Reversing the assembly motion also yields a small set of contact-aware intermediate poses. An insertion task receives an aligned pre-insertion pose; a screwing task receives poses along the thread at 90-degree rotational intervals. These goals remain sparse, because they describe success configurations instead of an engineered reward for every action. Finetuning starts from the play policy and uses Split and Aggregate Policy Gradients (SAPG).
The deployment pipeline uses the same CAD meshes with FoundationPose to track both the movable part and fixture in 6D. Policies run closed-loop at 60 Hz, pose tracking at 30 Hz, and domain randomization covers action latency plus delays and noise in current and goal poses.
What the Ablations Actually Say
The four ablations are a useful design guide for pretraining embodied policies.
- Object diversity: pretraining on 1,000 primitives produces more stable finetuning than 100 or 10 objects. Diversity includes inertial properties, not just geometry.
- 6D objective: translation-only play can learn lifting and transport but misses the orientation-control prior required for assembly. Rotation-only play transfers much better; full 6D goals are most consistent because they couple reorientation with translation.
- Trajectory diversity: freshly sampled goal trajectories outperform fixed banks of 10 or 100 trajectories. The policy gains broader coverage of object-pose transitions.
- Goal precision: a 10 cm success tolerance transfers poorly and 5 cm learns more slowly. The 1 cm objective teaches the fine object-pose control that tight clearance later demands.
Together, these results sharpen the paper’s thesis. Pretraining data is useful when its objective makes the desired control skill unavoidable. A large amount of play that permits a fixed grasp and arm-only transport is much less relevant to contact-rich assembly.
Results: Fast RL, Then Contact-Aware Behavior
The authors evaluate tight T-peg insertion, two stages of multi-part beam assembly, and screwing a furniture leg. Across the four contact-rich tasks, Play2Perfect reaches successful policies in roughly 2–5 hours of wall-clock RL. Scratch policies with either sparse rewards or hand-designed dense multi-stage rewards obtain no successful rollouts after 24 hours.
On an easier fixtured insertion variant, scratch training can eventually work, but dense-reward scratch needs more than 100 hours to approach perfect success; Play2Perfect reaches the same level in 4 hours, a reported 33× speed-up. The behavioral distinction matters: the dense-reward policy balances the peg with its thumb, then collapses under perturbation. At a 10 N perturbation it falls to about 20% success and reaches zero under stronger disturbances. The play-pretrained policy forms a multi-finger grasp and stays above 75% even at the largest tested perturbations.
Play alone is still insufficient for fine assembly. In simulation, the frozen play policy succeeds about 75% of the time at 40 mm insertion clearance and falls near zero by 4 mm. After assembly finetuning, Play2Perfect reports 95% at 4 mm, 92% at 1 mm, and 80% at 0.2 mm, including a tighter setting than its training range. The adaptation phase learns local contact search and corrective motions instead of treating contact as a disturbance.
Zero-Shot Sim-to-Real Transfer
Without real-world finetuning, the system achieves 10/10, 9/10, and 6/10 successes for tight insertion at 10 mm, 2 mm, and 0.5 mm clearance. Its two beam-assembly stages achieve 8/10 and 7/10, while the furniture-leg task reaches 7/10 for insertion and 5/10 for full screwing. Completion time increases as clearance tightens—from (6.8\pm1.5) s at 10 mm to (11.1\pm5.1) s at 0.5 mm—consistent with the policy taking more local alignment actions.
These results are encouraging, especially the 60% success rate at 0.5 mm clearance. They should still be read at the scale of the reported evaluation: ten hardware trials per condition, fixed fixture pose, and randomized initial part poses. The most common failures remain at the final contact-rich interaction, where occlusion degrades pose estimates and contact dynamics expose simulation-to-reality mismatch.
Takeaways and Limits
For dexterous robotics, Play2Perfect reframes play as RL pretraining, not a zero-shot replacement for task-specific control. Free-space play solves the broad skill-acquisition problem; sparse assembly finetuning learns the geometry- and contact-specific part. This division is a practical answer to sparse rewards when demonstrations, fixtures, and hand-engineered reward staging are costly.
The method also has boundaries. It learns short-horizon skills, while task sequencing, part selection, and goal poses are supplied externally. It finetunes per task or benchmark family and depends on CAD-based pose tracking; the policy does not directly observe the fixture or surrounding geometry beyond goal poses. Integrating scene perception, tactile sensing, recovery behavior, and a higher-level sequencing policy would be the natural route from individual assembly skills toward autonomous multi-part assembly.