{"id":50612116,"url":"https://github.com/PKU-EPIC/GraspVLA","last_synced_at":"2026-06-23T00:00:56.716Z","repository":{"id":291881814,"uuid":"978518470","full_name":"PKU-EPIC/GraspVLA","owner":"PKU-EPIC","description":"GraspVLA: a Grasping Foundation Model Pre-trained on Billion-scale Synthetic Action Data","archived":false,"fork":false,"pushed_at":"2025-07-25T16:03:33.000Z","size":4638,"stargazers_count":145,"open_issues_count":5,"forks_count":2,"subscribers_count":14,"default_branch":"main","last_synced_at":"2025-07-25T23:04:25.540Z","etag":null,"topics":[],"latest_commit_sha":null,"homepage":"","language":"Python","has_issues":true,"has_wiki":null,"has_pages":null,"mirror_url":null,"source_name":null,"license":null,"status":null,"scm":"git","pull_requests_enabled":true,"icon_url":"https://github.com/PKU-EPIC.png","metadata":{"files":{"readme":"README.md","changelog":null,"contributing":null,"funding":null,"license":null,"code_of_conduct":null,"threat_model":null,"audit":null,"citation":null,"codeowners":null,"security":null,"support":null,"governance":null,"roadmap":null,"authors":null,"dei":null,"publiccode":null,"codemeta":null,"zenodo":null}},"created_at":"2025-05-06T05:37:29.000Z","updated_at":"2025-07-25T16:03:36.000Z","dependencies_parsed_at":"2025-05-07T02:46:59.200Z","dependency_job_id":"83c3168b-c4a2-483b-acb7-b80b63aaab6c","html_url":"https://github.com/PKU-EPIC/GraspVLA","commit_stats":null,"previous_names":["pku-epic/graspvla"],"tags_count":0,"template":false,"template_full_name":null,"purl":"pkg:github/PKU-EPIC/GraspVLA","repository_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/PKU-EPIC%2FGraspVLA","tags_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/PKU-EPIC%2FGraspVLA/tags","releases_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/PKU-EPIC%2FGraspVLA/releases","manifests_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/PKU-EPIC%2FGraspVLA/manifests","owner_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners/PKU-EPIC","download_url":"https://codeload.github.com/PKU-EPIC/GraspVLA/tar.gz/refs/heads/main","sbom_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/PKU-EPIC%2FGraspVLA/sbom","scorecard":null,"host":{"name":"GitHub","url":"https://github.com","kind":"github","repositories_count":286080680,"owners_count":34669839,"icon_url":"https://github.com/github.png","version":null,"created_at":"2022-05-30T11:31:42.601Z","updated_at":"2026-05-26T15:22:16.424Z","status":"online","status_checked_at":"2026-06-22T02:00:06.391Z","response_time":106,"last_error":null,"robots_txt_status":"success","robots_txt_updated_at":"2025-07-24T06:49:26.215Z","robots_txt_url":"https://github.com/robots.txt","online":true,"can_crawl_api":true,"host_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub","repositories_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories","repository_names_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repository_names","owners_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners"}},"keywords":[],"created_at":"2026-06-06T05:00:21.752Z","updated_at":"2026-06-23T00:00:56.706Z","avatar_url":"https://github.com/PKU-EPIC.png","language":"Python","funding_links":[],"categories":["1. 机器人项目 | Robots","Robotic Grasping \u0026 Manipulation Perception"],"sub_categories":["1.1 具身智能与VLA大模型","Grasp Detection \u0026 Planning"],"readme":"# GraspVLA: a Grasping Foundation Model Pre-trained on Billion-scale Synthetic Action Data\n\n[![arXiv](https://img.shields.io/badge/arXiv-2505.03233-df2a2a.svg)](https://arxiv.org/pdf/2505.03233)\n[![Static Badge](https://img.shields.io/badge/Project-Page-a)](https://pku-epic.github.io/GraspVLA-web/)\n\n[**Model Server**](#model-server) | [**Simulation Playground**](#simulation-playground) | [**Real World Control Interface**](#real-world-control-interface)\n\nWe present a cost-effective pretraining paradigm for VLA models using only synthetic data, achieving direct sim-to-real transfer and strong zero-shot generalizability for robotic grasping. Key contributions include:\n\n- **SynGrasp-1B**: a billion-frame synthetic grasping dataset, spanning 240 object categories and 10,000+ objects.\n\n- **GraspVLA**: a VLA model pretrained on SynGrasp-1B that achieves zero-shot generalization to real-world grasping without fine-tuning.\n\n- **Unified CoT Framework**: GraspVLA integrates autoregressive perception and flow-matching-based action generation into a single reasoning process, enabling joint training on synthetic action data and internet-scale semantic data for open-vocabulary grasping.\n\n![teaser](./figs/teaser.jpg)\n\n## Latest Updates\n- [2025-12-29] 🎉🎉🎉 We release [StereoVLA](https://shengliangd.github.io/StereoVLA-Webpage/), a VLA model powered by stereo vision and supports flexible deployment with high tolerance to camera pose variations.\n- [2025-07-25] Release the GraspVLA model, [simulation playground](https://github.com/MiYanDoris/GraspVLA-playground) and [real world control interface](https://github.com/MiYanDoris/GraspVLA-real-world-controller).\n- [2025-07-19] Release the [supplementary material](https://arxiv.org/pdf/2505.03233).\n\n## Model Server\n\nPlease follow the steps below to start the model server. We provide the checkpoint of GraspVLA on both [huggingface](https://huggingface.co/shengliangd/GraspVLA) and [Baidu cloud](https://pan.baidu.com/s/1DOJbKrKzdBcEIrFQ_NcuLw?pwd=6666). GraspVLA achieves 200ms inference latency using ~9GB of GPU memory when running on a single NVIDIA RTX L40s GPU.\n\n### Step 1: Clone the Repository\n```bash\ngit clone https://github.com/PKU-EPIC/GraspVLA\ncd GraspVLA\n```\n\n### Step 2: Set Up Python Environment\nCreate and activate a conda environment with the required dependencies, for example:\n```bash\nconda create -n GraspVLA python=3.9.19\nconda activate GraspVLA\npip install -r requirements.txt\n```\n\n### Step 3: Download Model Weights\n\nIf you want to download from huggingface:\n```\npip install -U \"huggingface_hub\"\n# set HF_ENDPOINT if you encounter connection issues:\n# export HF_ENDPOINT=https://hf-mirror.com\nhf download shengliangd/GraspVLA\n```\nThe model weight will be placed at `~/.cache/huggingface/hub/models--shengliangd--GraspVLA/snapshots/f291eac1d3494c5c13c3d420af4e5bc987f23c3e/checkpoint/model.safetensors`.\n\n### Step 4: Launch the Model Server\nRun the model server with your desired configuration, for example:\n```bash\npython3 -u -m vla_network.scripts.serve --path you-path-to-model.safetensors --port 6666\n```\n\n#### Required arguments:\n\n- --path — Path to the model.safetensors file.\n\n- --port — Port number on which the server will listen for incoming requests.\n\n#### Optional arguments:\n\n- --compile: Enable model compilation (default: False). Speeds up inference (500ms → 200ms) but adds ~3 minutes to startup time. Recommended for large-scale evaluations (e.g., LIBERO benchmark).\n\nSuccess: The message `Started server on port \u003cPORT\u003e` indicates the server is ready.\n\n### Offline Test And Visualization\n\nTo quickly test GraspVLA without setting up a simulation or real-world environment, use the `offline_test` script. It runs offline inference on pre-recorded requests and compares results. The script saves a comparison image in `visualization`, with our reference output (left) and your model’s output (from the specified port) on the right.\n\n```bash\npython3 -u -m vla_network.scripts.offline_test --port \u003cPORT\u003e\n```\n\n### Repository Structure\n\nHigh-level overview of `vla_network` file-tree:\n\n+ `config/` - Contains basic configurations for our model.\n+ `data_preprocessing/` - Includes tools for preprocessing raw data into model-ready formats.\n+ `model/` - Code for defining and loading the main model structure.\n+ `scripts/` - Contains the `serve.py` file that starts a model server and the `offline_test.py` file that performs offline visualization.\n+ `type/` - Data type definitions used in our model.\n+ `utils/` - Contains some constants used in our model.\n\n## Simulation Playground\nWe provide a simulation playground for GraspVLA here: [GraspVLA-playground](https://github.com/MiYanDoris/GraspVLA-playground). This repository includes both the evaluation code used for GraspVLA in the [LIBERO](https://github.com/Lifelong-Robot-Learning/LIBERO) benchmark and an enhanced playground environment built on top of it. The playground provides an easy-to-use interface to evaluate GraspVLA across diverse objects, layouts, and environments.\n![playground](figs/playground.gif)\n\n## Real World Control Interface\nWe provide a [real-world control interface](https://github.com/MiYanDoris/GraspVLA-real-world-controller) for deploying GraspVLA in physical environments. This interface enables:\n\n- Zero-shot evaluation on real-world objects.\n\n- Both blocking and non-blocking control modes.\n\n- Real-time visualization of intermediate COT results, including 2D bounding boxes and 3D grasp poses.\n\n\u003cimg src=\"figs/real-world.png\" alt=\"camera_setup\" width=\"30%\" height=\"30%\" style=\"display: block; margin: 0 auto;\"\u003e\n\n## Citation\n\nIf you find this work useful, please cite:\n\n```bibtex\n@article{deng2025graspvla,\n    title={GraspVLA: a Grasping Foundation Model Pre-trained on Billion-scale Synthetic Action Data}, \n    author={Shengliang Deng and Mi Yan and Songlin Wei and Haixin Ma and Yuxin Yang and Jiayi Chen and Zhiqi Zhang and Taoyu Yang and Xuheng Zhang and Wenhao Zhang and Heming Cui and Zhizheng Zhang and He Wang},\n    year={2025},\n    eprint={2505.03233},\n    archivePrefix={arXiv},\n    primaryClass={cs.RO},\n    url={https://arxiv.org/abs/2505.03233}\n}\n```\n\n[![License](https://licensebuttons.net/l/by-nc/4.0/88x31.png)](LICENSE)\n","project_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2FPKU-EPIC%2FGraspVLA","html_url":"https://awesome.ecosyste.ms/projects/github.com%2FPKU-EPIC%2FGraspVLA","lists_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2FPKU-EPIC%2FGraspVLA/lists"}