[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"project-93980":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":14,"contributorsCount":14,"subscribersCount":14,"size":14,"stars1d":15,"stars7d":15,"stars30d":15,"stars90d":14,"forks30d":14,"starsTrendScore":16,"compositeScore":17,"rankGlobal":9,"rankLanguage":9,"license":18,"archived":19,"fork":19,"defaultBranch":20,"hasWiki":19,"hasPages":19,"topics":21,"createdAt":9,"pushedAt":9,"updatedAt":22,"readmeContent":23,"aiSummary":9,"trendingCount":14,"starSnapshotCount":14,"syncStatus":24,"lastSyncTime":25,"discoverSource":26},93980,"tau-0-vla","sii-research\u002Ftau-0-vla","sii-research","This repo is the official implementation of  \"τ0-VLA: a Hierarchical Robot Foundation Model with World-Model-Guided Test-Time Computation\".",null,"Python",200,9,1,0,37,111,85.2,"Apache License 2.0",false,"main",[],"2026-07-30 04:02:14","# τ₀-VLA: a Hierarchical Robot Foundation Model with World-Model-Guided Test-Time Computation\n\n\u003Cdiv id=\"top\" align=\"center\">\n\n![τ₀-VLA overview](assets\u002Foverview.png)\n\n\u003Ca href=\"https:\u002F\u002Ftau0-vla.github.io\u002F\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FProject_Website-tau0_VLA-blue\" height=\"25\" alt=\"Project Website\">\u003C\u002Fa>   \u003Ca href=\"https:\u002F\u002Ftau0-vla.github.io\u002Ftau0-vla.pdf\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FPaper-tau0_VLA-red\" height=\"25\" alt=\"Paper\">\u003C\u002Fa>   \u003Ca href=\"https:\u002F\u002Fhuggingface.co\u002Fsii-research\u002Ftau-0-vla\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FWeight-Hugging_Face-orange\" height=\"25\" alt=\"Model Weights\">\u003C\u002Fa>\n\n\u003C\u002Fdiv>\n\nThis repo is the official implementation of **τ₀-VLA: a Hierarchical Robot\nFoundation Model with World-Model-Guided Test-Time Computation**.\n\n## News\n\n- **[2026.07.27]** 🚀 We release the **τ₀-VLA** model\n  [Paper](https:\u002F\u002Ftau0-vla.github.io\u002Ftau0-vla.pdf),\n  [Project Website](https:\u002F\u002Ftau0-vla.github.io\u002F), and\n  [Hugging Face](https:\u002F\u002Fhuggingface.co\u002Fsii-research\u002Ftau-0-vla).\n\n## Overview\n\nτ₀-VLA is a hierarchical robot foundation model for long-horizon\nmanipulation. A memory-augmented high-level policy generates the next subtask\nand uses world-model-guided test-time computation to search over alternatives\nwhen additional reasoning is needed. A generalist low-level policy then\nexecutes the selected subtask across robot embodiments.\n\nThe low-level policy combines a Qwen3.5 vision-language backbone with a\nMixture-of-Transformers action expert trained through conditional flow\nmatching. It uses a unified 40-dimensional state\u002Faction space and was trained\non 40,115 hours of heterogeneous real-world robot data with multimodal\nco-training.\n\n![Hierarchical τ₀-VLA pipeline](assets\u002Fmethod.png)\n\n## Installation\n\nThe reference environment uses Python 3.11, CUDA 12.8, and PyTorch 2.7.1.\n\n```bash\ngit clone git@github.com:sii-research\u002Ftau-0-vla.git\ncd tau-0-vla\nbash scripts\u002Fsetup.sh\n```\n\n## Example data and post-training\n\n[`example_data\u002F`](example_data\u002FREADME.md) contains a small AgiBot World subset\nin LeRobot v3.0 format. The matching post-training recipe is under\n[`configs\u002Fexample_agibot_world_gong\u002F`](configs\u002Fexample_agibot_world_gong\u002FREADME.md).\n\n```bash\nbash scripts\u002Ftrain.sh configs\u002Fexample_agibot_world_gong\u002Ftrain.yaml \\\n    --model_name_or_path \u002Fpath\u002Fto\u002Ftau-0-vla-checkpoint\n```\n\nFor another dataset or robot, start from\n[`configs\u002F_template\u002F`](configs\u002F_template\u002FREADME.md) and\n[`src\u002Ftau0_vla\u002Fadapters\u002F_template\u002F`](src\u002Ftau0_vla\u002Fadapters\u002F_template\u002FREADME.md).\n\n## Serving and evaluation\n\nPublic v1 serving supports joint-control checkpoints only. Native EEF data may\nbe used for training, but EEF serving is not supported in this release.\n\nServe a post-trained joint-control checkpoint:\n\n```bash\npython -m deploy.server --model outputs\u002F\u003Crun_name>\n```\n\nRun open-loop evaluation:\n\n```bash\npython deploy\u002Fopenloop.py --ckpt outputs\u002F\u003Crun_name> --no-plot\n```\n\nSee [`deploy\u002F`](deploy\u002FREADME.md) for the payload and action-order contracts.\n\n## Repository layout\n\n```text\nsrc\u002Ftau0_vla\u002F\n├── adapters\u002F    embodiment-specific data layouts and deployment I\u002FO\n├── data\u002F        LeRobot loading, prompting, masking, and normalization\n├── models\u002F      Qwen3.5 backbone and flow-matching action expert\n├── trainer\u002F     post-training entry point\n├── vlm\u002F         multimodal collation and tokenization\n└── utils\u002F       logging and run specifications\nconfigs\u002F         reusable template and the AgiBot World example\ndeploy\u002F          policy server and open-loop evaluation\nexample_data\u002F    bundled AgiBot World subset\nscripts\u002F         setup, training, and normalization utilities\n```\n\nAdditional documentation:\n\n- [Dataset format](src\u002Ftau0_vla\u002Fdata\u002FDATASET_FORMAT.md)\n- [Data pipeline](src\u002Ftau0_vla\u002Fdata\u002FREADME.md)\n- [Robot adapters](src\u002Ftau0_vla\u002Fadapters\u002FREADME.md)\n\n## License\n\nCode and model weights are released under the [Apache License 2.0](LICENSE).\nThe bundled example data follows the license described in\n[`example_data\u002FREADME.md`](example_data\u002FREADME.md).\n",2,"2026-07-29 02:30:03","CREATED_QUERY"]