Kimodo

Kimodo is a controllable kinematic motion diffusion model developed by NVIDIA's Spatial Intelligence Lab (SIL). It is trained on 700 hours of optical motion capture data and generates high-quality 3D human motions that can be intuitively controlled through text prompts and a comprehensive suite of kinematic constraints.

Kimodo addresses the limitations of existing motion generation models by scaling up both dataset size and model capacity. While public mocap datasets are typically small, Kimodo leverages 700 hours of high-quality optical mocap data to produce more expressive, accurate, and generalizable motion outputs. The project is available on GitHub at research.nvidia.com/labs/sil/projects/kimodo.

Kimodo uses a diffusion-based generative approach, which means it starts from random noise and progressively refines it into realistic motion sequences. The key innovation lies in its carefully designed motion representation and two-stage denoiser architecture.

• Diffusion Model — Classifier-free guided diffusion for high-quality motion generation with diverse outputs
• Two-Stage Denoiser — Separates root (global translation) and body (joint rotations) prediction to minimize artifacts
• 700 Hours of Training Data — Trained on large-scale optical motion capture data for superior quality and generalization
• Flexible Constraint Conditioning — Supports multiple constraint types that can be combined in any configuration

Kimodo enables several powerful motion generation capabilities that can be combined for complex animation tasks:

• Text-to-Motion — Generate a wide range of behaviors from natural language descriptions like "a person walking confidently" or "someone tripping and recovering"
• Full-Body Keyframe Constraints — Specify exact poses at specific frames; the model generates smooth transitions between keyframes
• Sparse Joint Control — Constrain specific joints (hands, feet, elbows) to desired positions or rotations while letting the rest of the body move naturally
• 2D Waypoint Control — Guide the character's path through 2D waypoints on the ground plane
• Dense Path Following — Constrain the character to follow complex curved paths with natural pelvis motion

Kimodo supports a comprehensive suite of kinematic constraints that can be combined in any configuration:

Constraint Type	Description
Full-Body Keyframes	Specify complete body poses at particular frames for precise choreography
Joint Positions	Constrain the 3D position of specific joints (hands, feet, head, etc.)
Joint Rotations	Constrain the local rotation of specific joints for precise articulation
2D Waypoints	Guide the character's root translation through 2D points on the ground
Dense 2D Paths	Control the character to follow complex continuous paths

Kimodo's two-stage denoiser architecture decomposes motion prediction into separate root and body components. The first stage predicts the root trajectory (global position and rotation of the character's pelvis), while the second stage predicts the full body joint rotations conditioned on the root. This separation reduces motion artifacts like foot sliding and penetration that commonly plague single-stage models.

The model uses a smoothed root motion representation that enables following straight and curved paths closely while maintaining natural pelvis motion. Experiments on large-scale mocap data validate the design decisions and analyze how scaling dataset size and model size affect performance.

Kimodo has applications across multiple domains:

• Digital Humans — Generate realistic motion for digital characters (e.g., NVIDIA's SOMA) with detailed text prompts
• Robotics — Generate humanoid demonstration data more quickly and easily than teleoperation; motions can be exported for training physics-based policies via ProtoMotions and MuJoCo
• Animation & VFX — Rapid prototyping of character animations for films, games, and virtual productions
• Simulation — Generate diverse human behaviors for training AI systems in simulated environments
• Entertainment — Game development, virtual reality, and interactive experiences requiring realistic human motion

• Project Page — research.nvidia.com/labs/sil/projects/kimodo
• GitHub Code — Source code, motion authoring demo, and Python API at NVIDIA's GitHub
• Hugging Face Models — Pre-trained model weights available for download
• Technical Report — PDF paper detailing the architecture, training methodology, and experimental results
• Hugging Face Demo — Interactive demo for testing Kimodo's motion generation capabilities