Think about how a child learns to build with blocks.

They don't start by reading an assembly manual. They grasp, rotate, drop, re-grasp, and poke objects around — often with no specific goal beyond curiosity. Over time, those unstructured interactions become reusable skills: stable grasps, in-hand reorientation, controlled reaching. Only later do those skills compose into deliberate construction.

Precise robotic assembly has largely skipped this stage. Industry and research have pursued assembly head-on — with specialized fixtures, custom grippers, dense reward engineering, or painstaking teleoperation — while the underlying dexterity priors were left untrained.

In Play2Perfect: What Matters in Dexterous Play Pretraining for Precise Assembly?, researchers argue a simple but powerful thesis: before a robot can perfect precise assembly, it must first learn to play. They back it up with an RL framework that pretrains on diverse, task-agnostic play, then finetunes on sparse-reward assembly — achieving 33× higher sample efficiency than training from scratch, and zero-shot sim-to-real transfer on contact-rich tasks down to 0.5 mm clearance.

The Problem: Assembly Is Sparse, Contact-Rich, and Brutally Data-Hungry

Multi-fingered robot hands promise human-like speed and dexterity. Yet problems like tight insertion, multi-part assembly, and threaded screwing have remained stubbornly out of reach for learning-based methods. Three structural barriers explain why:

1. Sparse rewards make direct RL intractable

   Assembly success is binary and rare: the peg only "counts" when it clears the final millimeter of the hole; the screw only succeeds when threads engage. Random exploration almost never stumbles into reward — so policies trained from scratch on assembly tasks alone require astronomical interaction budgets.

2. Imitation learning is expensive and brittle

   Contact-rich assembly demands high-quality demonstrations of grasping, reorientation, alignment, and insertion — often with occluded views and delicate force profiles. Collecting enough expert trajectories for each assembly variant does not scale.

3. Prior work sidesteps dexterity instead of learning it

   Much progress has come from problem structuring: parallel-jaw grippers with chamfered fixtures, tool attachments, or environment jigs that reduce the problem to coarse positioning. That works in factories — but it does not produce general-purpose dexterous hands that can assemble in unstructured settings.

Play2Perfect asks a different question: What if we decouple "learning to manipulate" from "learning to assemble"?

The Solution: Play First, Perfect Later

Play2Perfect is a two-stage RL framework built on a clear division of labor:

The insight is developmental: play acquires broad motor competence cheaply; finetuning spends exploration budget only where it matters — the final contact-rich, sub-millimeter interactions that define assembly success.

How It Works: What Actually Matters in "Play"

The paper does not treat play as an undifferentiated data dump. It systematically studies which design choices in play pretraining transfer to assembly:

Object diversity Exposure to varied shapes, sizes, and materials during play builds grasp and reorientation policies that generalize when assembly introduces unfamiliar parts.

Training objective Goal-conditioned play — reach this pose, reorient to this configuration — produces structured behavior richer than random motor babbling, without requiring task labels.

Trajectory diversity Multiple paths to similar goals teach recovery: if one grasp fails, the policy has seen alternatives. This shows up dramatically at deployment (see below).

Goal precision How tightly play goals are specified affects how finely the prior resolves pose errors — critical when assembly tolerances shrink to sub-millimeter clearances.

Together, these choices produce a general manipulation prior that finetuning can adapt with far less data than RL from scratch — even when the scratch baseline receives dense, multi-stage shaped rewards.

Putting It to the Test: Three Assembly Benchmarks at 60 Hz

All deployment videos on the project page run at 1× speed with 60 Hz control — not sped-up sim footage. Tasks include:

Tight Insertion

Peg-in-hole with only 0.5 mm contact clearance — a tolerance where visual misalignment of a fraction of a degree causes jamming. Play2Perfect achieves 60% zero-shot real-world success without real-world training.

Screwing

Thread engagement demands controlled rotation, axial force, and compliance — the same failure mode that breaks light bulbs and strips screws when done wrong. Success exceeds 50% on long-horizon screwing tasks.

Multi-Part Assembly (Step 1 & 2)

Long-horizon sequencing: grasp part A, reorient, align, insert; then repeat for part B. Each stage compounds error — making the play-learned reorientation and pose-reaching priors essential.

These are exactly the tasks where anthropomorphic multi-finger hands earn their keep: parallel grippers cannot reorient a peg in-hand; suction cups cannot thread a screw. You need the dexterity of a human hand — or its robotic equivalent.

Recovery Behavior: The Hidden Benefit of Play

One of the most compelling results is not a success-rate number — it is what happens after failure.

Play2Perfect policies act closed-loop: when an insertion misaligns or a grasp slips, the robot does not freeze or abort. It continues acting, retrying and adjusting until the task completes. This recovery behavior emerges from play pretraining's trajectory diversity — the policy has seen suboptimal contact states before and learned to escape them.

For industrial deployment, this matters as much as first-attempt success rate: a robot that retries intelligently finishes jobs that open-loop trajectories abandon.

The Results: 33× Sample Efficiency, Zero-Shot Sim-to-Real

The sim-to-real result is particularly notable: policies trained entirely in simulation deploy on real hardware without additional real-world data collection or finetuning. This aligns with the broader trend Sharpa has highlighted in research like SimToolReal and Tacmap — simulation diversity + anthropomorphic hardware fidelity can close the gap that once demanded months of teleoperation.

Why This Matters for Sharpa Wave

Play2Perfect validates a hardware–software co-design principle central to Sharpa's platform: you cannot shortcut dexterity with fixtures if you want general-purpose manipulation.

1:1 anthropomorphic kinematics (22 DoF) Play pretraining learns in-hand reorientation and multi-finger pose control — capabilities that require a hand with human-like degrees of freedom. Sharpa Wave provides 22 active DoF across thumb, fingers, and wrist with human-scale proportions (palm width-to-length ratio ≈ 0.618), enabling the same grasp families and in-hand rotations that play pretraining discovers in simulation.

High-frequency closed-loop control (60 Hz)

Assembly at 0.5 mm clearance cannot be executed open-loop. Wave supports >4 Hz full-gesture speed with >20 N fingertip force — the mechanical bandwidth to execute and recover from contact at the rates Play2Perfect demonstrates.

Sim-to-real with Isaac Sim assets Sharpa Wave ships with Isaac Sim URDF/USD simulation models integrated into the NVIDIA Isaac GR00T Reference Humanoid workflow. Play2Perfect's zero-shot transfer is exactly the outcome this toolchain is built to enable: train play priors and assembly skills in sim, deploy on real Wave hands with minimal gap.

Tactile-ready for the last millimeter Play2Perfect's finetuning stage focuses exploration on contact-rich insertion — where vision fails and touch dominates. Wave's Dynamic Tactile Array (DTA) with >1,000 pixels per fingertip provides the sensing substrate for future work combining play priors with tactile-reactive finetuning — the direction explored in SaTA and T-Rex.

Task overlap with Sharpa's demonstrated capabilities

Screwing, tight insertion, and multi-part assembly are the same task class Sharpa North demonstrated at CES 2026 (card dealing at ~2 mm precision, 30+ step handicraft sequences) and that SaTA evaluated (light-bulb threading, USB mating at sub-millimeter tolerance). Play2Perfect shows a scalable training recipe — play then perfect — for acquiring these skills without task-specific fixtures.

Consistency across the research stack

Play2Perfect complements other Sharpa-featured research on the same problem axis:

• SimToolReal — procedural sim pretraining for zero-shot tool manipulation
• EgoScale — human video pretraining on Sharpa Wave's 22-DoF action space
• SaTA — spatially grounded tactile policies for USB and screwing

Together, they sketch a convergent picture: diverse pretraining (play, human video, procedural sim) + anthropomorphic tactile hardware + sparse finetuning on real tasks = dexterous assembly without factory fixtures.

From Play to Product: Manufacturing Time, One Insertion at a Time

Sharpa's mission — "We manufacture time by making robots useful" — depends on robots that handle contact-rich work autonomously, not just approach it. Play2Perfect offers a practical curriculum for getting there:

1. Pretrain through play on diverse objects — cheap, scalable, no task labels
2. Finetune on assembly with sparse rewards — sample-efficient because priors already exist
3. Deploy zero-shot from simulation — no months of real-world data collection

For OEMs and researchers building on Sharpa Wave, the takeaway is clear: invest in general dexterity before task-specific assembly. The hand that learns to play is the hand that learns to build.