--- title: "TCP4D: A Unified Tool-Center-Point Tracking and Generation Model for Robot Manipulation" emoji: 🦾 colorFrom: blue colorTo: indigo sdk: gradio sdk_version: 6.17.3 python_version: '3.12' app_file: app.py models: - HarrisonPENG/TCP4D-stage1-DROID preload_from_hub: - HarrisonPENG/TCP4D-stage1-DROID model.safetensors,config.json 021cff1437c5734b0c49b3935ed355862c89875c fullWidth: true short_description: Unified TCP tracking and generation for robot manipulation. tags: - robotics - point-cloud - droid - zerogpu --- # TCP4D: A Unified Tool-Center-Point Tracking and Generation Model for Robot Manipulation *TCP4D reconstructs scenes, tracks TCP poses and gripper states, predicts future TCP trajectories, and supports cross-embodiment data augmentation.* [GitHub](https://github.com/Livioni/TCP4D) Select an episode and camera, click the starting-frame TCP, then choose **Track TCP**. Explore the reconstructed point cloud and TCP poses in the embedded Viser viewer. Frames are preloaded for playback without rebuilding the scene or resetting the camera. Playback, geometry, TCP controls, and PLY export are inside the viewer. The default confidence filter hides the lowest-confidence 10% of scene points. Download the predictions, RGB overlay video, or current point cloud. Uses [HarrisonPENG/TCP4D-stage1-DROID](https://huggingface.co/HarrisonPENG/TCP4D-stage1-DROID) and the inference code from [TCP4D](https://github.com/Livioni/TCP4D). The checkpoint revision is pinned for reproducible inference. GPU work runs on ZeroGPU. Choose a DROID or DROID-compatible RoboLab example, or upload one episode as a ZIP. Local examples are discovered from `datasets/` and `assets/samples/droid//` and included in the deployment bundle. Required: `metadata.json`, `images//`, and `intrinsic/.npy`. RGB is 320x180 with contiguous zero-based frame names. Optional `TCP//state.npy` enables starting-frame GT projection; subsequent windows use predicted TCP queries. See `hf_demo/README.md` for the sample format. The original TCP4D LICENSE and NOTICE are included. Bundled episodes retain their original metadata and available upstream dataset attribution. ## References * [Model Paper](https://huggingface.co/papers/2611.23143)