[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"project-92906":3},{"id":4,"name":5,"fullName":6,"owner":7,"repo":5,"description":8,"homepage":9,"htmlUrl":9,"language":10,"languages":9,"totalLinesOfCode":9,"stars":11,"forks":12,"watchers":13,"openIssues":13,"contributorsCount":14,"subscribersCount":14,"size":14,"stars1d":14,"stars7d":14,"stars30d":14,"stars90d":14,"forks30d":14,"starsTrendScore":14,"compositeScore":15,"rankGlobal":9,"rankLanguage":9,"license":16,"archived":17,"fork":17,"defaultBranch":18,"hasWiki":19,"hasPages":17,"topics":20,"createdAt":9,"pushedAt":9,"updatedAt":21,"readmeContent":22,"aiSummary":23,"trendingCount":14,"starSnapshotCount":14,"syncStatus":24,"lastSyncTime":25,"discoverSource":26},92906,"play2perfect","kushal2000\u002Fplay2perfect","kushal2000","Official code release for Play2Perfect: a 2-stage RL pipeline for contact-rich assembly (arXiv 2606.26428)",null,"Python",53,5,1,0,42.33,"MIT License",false,"main",true,[],"2026-07-22 04:02:07","# Play2Perfect\n\n### What Matters in Dexterous Play Pretraining for Precise Assembly?\n\nOfficial code release for **Play2Perfect**, a 2-stage reinforcement-learning pipeline for\ncontact-rich robotic assembly.\n\n📄 **Paper:** https:\u002F\u002Farxiv.org\u002Fabs\u002F2606.26428\n\nhttps:\u002F\u002Fgithub.com\u002Fuser-attachments\u002Fassets\u002F6bfafdbc-41f4-472a-aeaa-444f7e50076b\n\n## Overview\n\nPlay2Perfect learns precise, contact-rich assembly in two stages:\n\n1. **Play** — pretrain a policy by playing with random objects in free space.\n2. **Finetune** — finetune the play policy on a contact-rich assembly task.\n\nThis repository provides:\n\n- **Training** code for both stages in Isaac Sim (Isaac Lab).\n- **Interactive (viser) evaluation** for four assembly tasks using our released checkpoints.\n- A **minimal sim-to-real deployment** reference.\n\n> This release targets **Isaac Sim only**.\n\n## Project structure\n\n```\nplay2perfect\u002F\n├── isaacsimenvs\u002F          # Training: Stage-1 play env + Stage-2 precise-assembly env\n│   ├── train.py           # Single training entry point (both stages)\n│   ├── cfg\u002F               # Hydra task + train configs\n│   └── tasks\u002F{play,precise_assembly}\u002F\n├── evaluation\u002F         # The 4 assembly tasks: problem registry + viser evaluation\n│   ├── problems\u002F          # peg-in-hole, fabrica beam parts 0 & 2, furniture-bench\n│   └── eval_isaacsim.py   # Interactive viser evaluation\n├── rl_games\u002F              # Vendored RL library (PPO + SAPG)\n├── deployment\u002F            # Minimal sim-to-real reference\n├── assets\u002Furdf\u002F           # Robot + task assets\n└── docs\u002F                  # Installation + deployment guides\n```\n\n## The four assembly tasks\n\n| Task | `problem` key |\n|------|----------------|\n| Tight insertion (L-peg, 0.5 mm tolerance) | `tight_insertion` |\n| Beam assembly — step 1 | `beam_assembly_step1` |\n| Beam assembly — step 2 | `beam_assembly_step2` |\n| Screwing (furniture leg) | `screwing` |\n\n## Installation\n\nSee [docs\u002Fisaacsim_installation.md](docs\u002Fisaacsim_installation.md) for the full Isaac Sim \u002F\nIsaac Lab setup (Python 3.11, `.venv_isaacsim`).\n\n## Quick start\n\n### 1. Download checkpoints\n\n```bash\n# Stage-1 play policy + the 4 finetuned assembly policies\npython download_checkpoints.py\n```\n\n### 2. Interactive evaluation (viser)\n\n```bash\n# One problem from a single checkpoint\npython evaluation\u002Feval_isaacsim.py \\\n  --checkpoint-path pretrained_assembly\u002Ftight_insertion\u002Fmodel.pth \\\n  --problem tight_insertion \\\n  --port 8044\n\n# ...or load all four downloaded policies and switch between them in the GUI\npython evaluation\u002Feval_isaacsim.py --policies-dir pretrained_assembly --port 8044\n```\n\nOpen `http:\u002F\u002Flocalhost:8044` in a browser to watch the policy roll out the insertion.\n\n### Offline success-rate evaluation\n\nHeadless, batched eval that scores each problem over many parallel envs (one\ni.i.d. first-episode trial per env) and prints a success-rate table:\n\n```bash\npython evaluation\u002Feval_offline.py --policies-dir pretrained_assembly\n```\n\n### 3. Train Stage 1 (play)\n\n```bash\npython isaacsimenvs\u002Ftrain.py \\\n  --task Isaacsimenvs-Play-Direct-v0 \\\n  --agent rl_games_cfg_entry_point \\\n  --headless\n```\n\n### 4. Finetune Stage 2 (precise assembly)\n\n```bash\npython isaacsimenvs\u002Ftrain.py \\\n  --task Isaacsimenvs-PreciseAssembly-Direct-v0 \\\n  --agent rl_games_cfg_entry_point \\\n  --checkpoint pretrained_policy\u002Fmodel.pth \\\n  env.precise_assembly.problem=tight_insertion \\\n  --headless\n```\n\n## Deployment\n\nSee [docs\u002Fdeployment.md](docs\u002Fdeployment.md) for the minimal sim-to-real reference.\n\n## Formatting\n\n```bash\n.\u002Fformat.sh   # ruff over the Python packages + xmllint over URDFs\n```\n\n## Acknowledgements\n\nThis codebase builds on excellent prior work:\n\n- [**rl_games**](https:\u002F\u002Fgithub.com\u002FDenys88\u002Frl_games) — the high-throughput RL library our\n  training loop is built on (vendored here).\n- [**SAPG**](https:\u002F\u002Fsapg-rl.github.io\u002F) — Split and Aggregate Policy Gradients, the\n  large-scale PPO variant we use for training.\n- [**SimToolReal**](https:\u002F\u002Fgithub.com\u002Ftylerlum\u002Fsimtoolreal) — our prior work on\n  object-centric tool manipulation, which the Stage-1 \"play\" pretraining is based on.\n\n## Citation\n\n```bibtex\n@misc{lum2026play2perfectmattersdexterousplay,\n      title={Play2Perfect: What Matters in Dexterous Play Pretraining for Precise Assembly?},\n      author={Tyler Ga Wei Lum and Kushal Kedia and C. Karen Liu and Jeannette Bohg},\n      year={2026},\n      eprint={2606.26428},\n      archivePrefix={arXiv},\n      primaryClass={cs.RO},\n      url={https:\u002F\u002Farxiv.org\u002Fabs\u002F2606.26428},\n}\n```\n\n## Contact\n\n- Kushal Kedia — kk837@cornell.edu\n- Tyler Lum — tylerlum@stanford.edu\n","Play2Perfect 是一个面向接触密集型机器人装配任务的两阶段强化学习（RL）框架。它通过第一阶段‘自由交互预训练’（Play）在无约束环境中学习灵巧操作通用技能，再第二阶段‘任务微调’（Finetune）针对具体装配任务（如紧配合插入、梁结构组装、螺丝拧紧等）进行精细化优化。项目基于 Isaac Sim\u002FIsaac Lab 构建，提供完整训练、可视化交互评估（viser）、离线成功率评测及轻量级仿真到真实部署参考。适用于需要高精度接触感知与动态力控的工业装配、机器人灵巧操作研究等场景。",2,"2026-07-11 02:30:16","CREATED_QUERY"]