--- source: 'https://howaiworks.ai/blog/nvidia-kimodo-3d-motion-generation' section: blog title: 'NVIDIA Kimodo: AI-Powered 3D Motion Generation' description: >- NVIDIA releases Kimodo, a diffusion-based generative model for realistic 3D motion, supporting diverse skeletons including SMPL-X and Unitree G1. date: '2026-03-24' author: HowAIWorks Team tags: - NVIDIA - AI - Robotics - 3D Animation - Diffusion Models - Kimodo - SMPL-X - Unitree G1 readingTime: 3 minutes --- # NVIDIA Kimodo: AI-Powered 3D Motion Generation > NVIDIA releases Kimodo, a diffusion-based generative model for realistic 3D motion, supporting diverse skeletons including SMPL-X and Unitree G1. ## Introduction NVIDIA has recently unveiled **Kimodo**, a state-of-the- art generative model designed to revolutionize how we create 3D motions for both digital avatars and physical robots. Based on [diffusion architecture](https://howaiworks.ai/blog/radical-numerics-rnd1-diffusion-model), Kimodo allows creators and engineers to generate highly realistic and precisely controlled movements from simple inputs like text descriptions or sparse keyposes. As the demand for lifelike animation in gaming, metaverse applications, and robotics continues to grow, Kimodo addresses the challenge of producing fluid, natural-looking motion that adheres to specific constraints. Whether it's a character walking across a room to sit on a chair or a humanoid robot performing complex tasks, Kimodo provides a unified framework for high-fidelity motion synthesis. ## Core Features and Capabilities Kimodo stands out due to its versatile control system and broad skeletal support. Unlike traditional motion synthesis tools that might rely solely on text-to-motion, Kimodo integrates multiple layers of control: - **Multimodal Inputs**: Supports text prompts for high-level descriptions and specific keyposes for precise structural control. - **Detailed Joint Control**: Allows for the specification of individual joint positions and rotations, enabling fine-tuning of limb placement. - **Path and Trajectory Control**: Creators can define 2D paths for characters to follow, ensuring spatial accuracy in motion. - **Real-time Editing**: The project includes an interactive web demo with a timeline editor for intuitive motion refinement. ## Model Variants and Training NVIDIA has released five distinct variants of the Kimodo model, optimized for different use cases and skeletal structures: - **Human Avatars (SOMA & SMPL-X)**: These skeletons are widely used in the animation and computer vision communities. The SMPL-X variant provides detailed control over body and hand movements. - **Robot Skeletons (Unitree G1)**: Specifically designed for the Unitree G1 humanoid robot, bridge the gap between digital animation and physical robotics. The models were trained on substantial datasets, including **Bones Rigplay 1** (700 hours of motion data) for production-ready performance, and **BONES-SEED** (288 hours) for benchmarking purposes. ## Integration with the NVIDIA Ecosystem Kimodo is not just a standalone generation tool; it is deeply integrated into [NVIDIA's broader robotics](https://howaiworks.ai/blog/uma-launches-physical-ai-robotics-2025) and simulation stack: 1. **ProtoMotions**: This integration allows developers to take generated motions and use them to train physically-accurate control policies within GPU-accelerated simulations. 2. **General Motion Retargeting**: Using this feature, motions generated for the SMPL-X skeleton can be seamlessly transferred to a wide variety of other robotic structures, making Kimodo a versatile tool for cross-platform development. ## Licensing and Availability NVIDIA has committed to open-source principles for Kimodo, though the licensing varies by component: - **Project Code**: Released under the **Apache 2.0** license, allowing for broad commercial and research use. - **Model Weights**: Most models are available under the **NVIDIA Open Model License**. However, the **SMPL-X** variant is governed by the more restrictive **NVIDIA R&D Model License**, which limits its use to non-commercial research. ## Conclusion The release of Kimodo represents a significant leap forward in AI-driven motion generation. By combining the flexibility of diffusion models with the precision of skeletal control, NVIDIA has provided a powerful tool for animators and roboticists alike. As these models continue to evolve, we can expect even more seamless integration between human-like digital animation and the physical capabilities of next-generation humanoid robots. ## Sources - [Original Telegram Announcement](https://t.me/ai_machinelearning_big_data/9727) - Official Kimodo Github Repository - NVIDIA Technical Report (Likely linked in the repo) ## Frequently Asked Questions ### What is NVIDIA Kimodo? Kimodo is a generative model based on diffusion that creates realistic 3D movements for human and robotic skeletons using various control signals like text and keyposes. ### Which skeletons does Kimodo support? Currently, Kimodo supports three types of skeletons: SOMA and SMPL-X for human avatars, and Unitree G1 for humanoid robotics. ### What controls can be used with Kimodo? Users can control motion generation using text prompts, full-body keyposes, joint positions, 2D paths, and limb-specific control points. ### What are the hardware requirements for running Kimodo? Generating motions with Kimodo requires approximately 17 GB of VRAM, typical of high-end GPUs like the RTX 3090, RTX 4090, or NVIDIA A100. --- Source: https://howaiworks.ai/blog/nvidia-kimodo-3d-motion-generation — HowAIWorks.ai