Play2Perfect

What Matters in Dexterous Play Pretraining for Precise Assembly?

Official code release for Play2Perfect, a 2-stage reinforcement-learning pipeline for contact-rich robotic assembly.

📄 Paper: https://arxiv.org/abs/2606.26428

highlight_reel.mp4

Overview

Play2Perfect learns precise, contact-rich assembly in two stages:

1. Play — pretrain a policy by playing with random objects in free space.
2. Finetune — finetune the play policy on a contact-rich assembly task.

This repository provides:

• Training code for both stages in Isaac Sim (Isaac Lab).
• Interactive (viser) evaluation for four assembly tasks using our released checkpoints.
• A minimal sim-to-real deployment reference.

This release targets Isaac Sim only.

Project structure

play2perfect/
├── isaacsimenvs/          # Training: Stage-1 play env + Stage-2 precise-assembly env
│   ├── train.py           # Single training entry point (both stages)
│   ├── cfg/               # Hydra task + train configs
│   └── tasks/{play,precise_assembly}/
├── evaluation/         # The 4 assembly tasks: problem registry + viser evaluation
│   ├── problems/          # peg-in-hole, fabrica beam parts 0 & 2, furniture-bench
│   └── eval_isaacsim.py   # Interactive viser evaluation
├── rl_games/              # Vendored RL library (PPO + SAPG)
├── deployment/            # Minimal sim-to-real reference
├── assets/urdf/           # Robot + task assets
└── docs/                  # Installation + deployment guides

The four assembly tasks

Task	problem key
Tight insertion (L-peg, 0.5 mm tolerance)	tight_insertion
Beam assembly — step 1	beam_assembly_step1
Beam assembly — step 2	beam_assembly_step2
Screwing (furniture leg)	screwing

Installation

See docs/isaacsim_installation.md for the full Isaac Sim / Isaac Lab setup (Python 3.11, .venv_isaacsim).

Quick start

1. Download checkpoints

# Stage-1 play policy + the 4 finetuned assembly policies
python download_checkpoints.py

2. Interactive evaluation (viser)

# One problem from a single checkpoint
python evaluation/eval_isaacsim.py \
  --checkpoint-path pretrained_assembly/tight_insertion/model.pth \
  --problem tight_insertion \
  --port 8044

# ...or load all four downloaded policies and switch between them in the GUI
python evaluation/eval_isaacsim.py --policies-dir pretrained_assembly --port 8044

Open http://localhost:8044 in a browser to watch the policy roll out the insertion.

Offline success-rate evaluation

Headless, batched eval that scores each problem over many parallel envs (one i.i.d. first-episode trial per env) and prints a success-rate table:

python evaluation/eval_offline.py --policies-dir pretrained_assembly

3. Train Stage 1 (play)

python isaacsimenvs/train.py \
  --task Isaacsimenvs-Play-Direct-v0 \
  --agent rl_games_cfg_entry_point \
  --headless

4. Finetune Stage 2 (precise assembly)

python isaacsimenvs/train.py \
  --task Isaacsimenvs-PreciseAssembly-Direct-v0 \
  --agent rl_games_cfg_entry_point \
  --checkpoint pretrained_policy/model.pth \
  env.precise_assembly.problem=tight_insertion \
  --headless

Deployment

See docs/deployment.md for the minimal sim-to-real reference.

Formatting

./format.sh   # ruff over the Python packages + xmllint over URDFs

Acknowledgements

This codebase builds on excellent prior work:

• rl_games — the high-throughput RL library our training loop is built on (vendored here).
• SAPG — Split and Aggregate Policy Gradients, the large-scale PPO variant we use for training.
• SimToolReal — our prior work on object-centric tool manipulation, which the Stage-1 "play" pretraining is based on.

Citation

@misc{lum2026play2perfectmattersdexterousplay,
      title={Play2Perfect: What Matters in Dexterous Play Pretraining for Precise Assembly?},
      author={Tyler Ga Wei Lum and Kushal Kedia and C. Karen Liu and Jeannette Bohg},
      year={2026},
      eprint={2606.26428},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2606.26428},
}

Contact

• Kushal Kedia — kk837@cornell.edu
• Tyler Lum — tylerlum@stanford.edu