nv-tlabs/kimodo

Official implementation of Kimodo, a kinematic motion diffusion model for high-quality human(oid) motion generation.

What it solves

Kimodo addresses the challenge of generating high-quality, controllable 3D human and robot motions. It allows users to create specific animations by combining natural language text prompts with precise kinematic constraints, such as joint positions, end-effector targets (hands/feet), and ground-plane paths.

How it works

Kimodo is a kinematic motion diffusion model trained on a large-scale motion capture dataset (700 hours of Rigplay data). It functions by taking a text description and optional spatial constraints as input, then iteratively refining a motion sequence through a diffusion process. It supports multiple skeleton types, including SOMA (humanoid), Unitree G1, and SMPL-X, and provides a web-based interactive timeline editor for users to author and preview these motions.

Who it’s for

This project is designed for researchers and developers working in robotics, character animation, and physical AI who need to generate realistic motion data for simulation or downstream tasks like training humanoid control policies.

Highlights

  • Multi-modal Control: Combines text-to-motion generation with fine-grained kinematic constraints like keyframes and 2D waypoints.
  • Interactive Authoring: Includes a local web-based GUI for timeline-based motion editing and real-time 3D visualization.
  • Robot Compatibility: Supports direct export to formats compatible with MuJoCo and other robotics frameworks like ProtoMotions.
  • Standardized Benchmarking: Provides a comprehensive evaluation suite and test cases to measure motion quality and constraint-following accuracy.
  • Flexible Inference: Offers both a high-level CLI for batch generation and a low-level Python API for custom integration.

Related

  • Project
    Tencent-Hunyuan/HY-Motion-1.0

    A series of billion-parameter text-to-3D human motion generation models that create skeleton-based animations from text prompts using Diffusion Transformers and Flow Matching.

  • Project
    aimagelab/multimodal-garment-designer

    A multimodal fashion image editing tool based on latent diffusion models that generates human-centric images from text, body poses, and garment sketches.

  • Project
    NVlabs/ProtoMotions

    ProtoMotions 3 is an Apache‑2.0‑licensed, GPU‑accelerated framework for large‑scale motion learning, retargeting, and sim‑to‑real deployment of humanoid robots. It supports multi‑GPU training on datasets such as AMASS and BONES, one‑command retargeting via PyRoki, cross‑simulator testing (Isaac Gym, Newton, MuJoCo), and export of ONNX policies for zero‑shot deployment on real hardware (e.g., Unitree G1).

  • Project
    AIGAnimation/CAMDM

    A real-time character animation system using diffusion models to generate high-quality, diverse movements based on user control signals and style presets.

  • Project
    zju3dv/MotionStreamer

    A streaming human motion generation system that uses a diffusion-based autoregressive model in a causal latent space to create continuous animations from text descriptions.