Kimodo is a kinematic motion diffusion model that generates high-quality 3D human and robot motions from text prompts and kinematic constraints.
The tool addresses motion generation by training a diffusion model on a large-scale commercially-friendly motion capture dataset. It accepts text descriptions alongside an extensive set of kinematic controls including full-body pose keyframes, end-effector positions and rotations, 2D paths, and 2D waypoints. The model operates on multiple skeleton formats and produces physically plausible motion sequences that respect both textual intent and geometric constraints.
Kimodo suits projects requiring controllable motion synthesis where both semantic guidance and precise spatial constraints matter. The tool provides inference code and a command-line interface for both human and robot skeleton generation, plus an interactive timeline-based demo for authoring motions. A motion generation benchmark with test cases and evaluation code is included, built on a public dataset with fine-grained temporal text annotations. This makes it useful for animation pipelines, robotics simulation, and motion research where reproducible evaluation is needed.
The project maintains multiple model variants trained on different skeletons and datasets, all supporting the same text and kinematic control interface. Development has included bug fixes to metric calculation in the benchmark, support for reduced-memory GPU configurations, and improvements to multi-prompt generation workflows. The codebase has evolved through skeleton format updates and expanded model availability across different skeleton representations.
