Kimodo adds local AI humanoid animation generation to an existing Unity project. It supports prompt-driven motion, pose and end-effector constraints, analysis, baking, retargeting, Animator content, Timeline authoring, and runtime motion playback.
- Runs from a project-owned local runtime.
- Supports Windows, macOS, and Linux; CUDA is the primary acceleration path and CPU fallback is available.
- Keeps animation generation and assets inside the Unity project workflow.
- Unity 2022.3 or newer
- An existing Unity project
- A character with a valid Humanoid Animator for humanoid workflows
- At least 8 GB memory and sufficient disk space for the selected models
In Unity Package Manager, choose Add package from git URL and enter:
https://github.com/OneYoungMean/KimodoUnityBridge.git
For a manifest dependency, use the same Git URL:
"com.unity.kimodo_unity_motion_tools": "https://github.com/OneYoungMean/KimodoUnityBridge.git"- Import the package's Light Sample from Unity Package Manager, or open the FullDemo.
- Open its Timeline and select a Kimodo clip bound to a character.
- Enter a motion prompt and choose Generate & Bake.
- Wait for the project-local runtime and model to become ready, then play the Timeline.
Runtime diagnostics are written to NvlabKimodoQuickServer~/log/setup.log and NvlabKimodoQuickServer~/log/bridge_server.log in the Unity project.
The QuickServer TCP generate command accepts output_mp4: true to save an MP4 render in the QuickServer directory while preserving the selected motion output format. The JSON response includes the absolute mp4_path; the binary payload remains only the selected motion format. The launcher discovers C:\nvlab\UniMate beside the repositories by default; set KIMODO_UNIMATE_ROOT to override it.
Read SKILL.md for purpose, installation, and documentation navigation; the animation workflow, task routing, tool coordination, and shared evidence rules are in tools/common.md. server.install and animation.generate are asynchronous and return a request_id; poll either with generation.get. Command definitions are kept in Command/help.json.
The asynchronous pose.fromvideo and pose.frompicture commands return SMPL-X pose arrays through generation.get. QuickServer bundles the Mixamo LLM Mocap pipeline and its 3.54 MB SMPLX_NEUTRAL_LITE_V437.npz asset for pose-only analysis; the full SMPL-X model is not required for these commands. Override the bundled source with KIMODO_MOCAP_ROOT, or set KIMODO_MOCAP_PYTHON to the GVHMR Python executable. Picture analysis estimates one pose from a short repeated still-image sequence.
Repository maintenance rules and development notes are in development/. The root AGENTS.md is only the repository-tool entry point and links to the detailed maintenance rules.
The documentation ownership is:
README.md/README.zh-CN.md: user-facing package documentation.SKILL.md: purpose, installation, and documentation navigation.tools/*.md: cooperating capability tools and their task procedures.Command/help.json: command definitions and parameter schema.development/: developer maintenance, compatibility, plans, and handoff notes.