Kimodo: Controllable Kinematic Motion Diffusion at Scale
Description:
Kimodo (Kinematic Motion Diffusion) generates three-dimensional (3D) skeletal body animations given a text prompt and/or constraints on the motion like full-body poses, end-effector joint positions, paths, and waypoints to follow.
The Kimodo model family includes models trained on different skeletons and datasets:
- Kimodo-SOMA-RP
- Trained on the 30-joint SOMA skeleton with the proprietary Bones Rigplay dataset.
- Kimodo-SOMA-SEED
- Trained on the 30-joint SOMA skeleton with the open Bones-SEED dataset.
- Kimodo-G1-RP
- Trained on the proprietary Bones Rigplay dataset retargeted to the 34-joint Unitree G1 robot skeleton.
- Kimodo-G1-SEED
- Trained on the open Bones-SEED dataset retargeted to the 34-joint Unitree G1 robot skeleton.
- Kimodo-SMPLX-RP
- Trained on the proprietary Bones Rigplay dataset retargeted to the 22-joint SMPLX-body skeleton.
This release pertains to Kimodo-SOMA-RP-v1.1. This model is ready for commercial use.
Changes In v1.1
Kimodo-SOMA-RP-v1.1 contains minor improvements over Kimodo-SOMA-RP-v1 and makes the model compatible with the
Kimodo Motion Generation Benchmark:
- Updated the training split to not overlap with the test splits of the Kimodo benchmark, making it comparable with other models trained on BONES-SEED
- As a result of the updated splits, the training set is slightly larger than in v1, resulting in more diversity seen during training
- Further data cleaning performed to remove problematic motions with wrist and shoulder twist artifacts
- Improved training stability
License:
This model is released under the
NVIDIA Open Model License.
Deployment Geography:
Global
Use Case:
The model is intended for users with any level of animation experience to create 3D human motion data for their application. This may include:
- Demonstrations for humanoid robots
- Digital human motion for digital twin and industrial simulations
- Digital human motion for synthetic data
- Animations for game and media development
Release Date:
Github [04/10/2026] via
link
HuggingFace [04/10/2026] via
link References:
Model Architecture:
Architecture Type: Diffusion Model
Network Architecture: Novel Two-Stage Transformer
Model Size: 282 M parameters
Inputs:
Input Types: Text, Duration (Num Frames), Pose Constraints
Input Formats:
- Text: String
- Duration: Integer
- Pose Constraints: Matrix
Input Parameters:
- Text: One-Dimensional (1D)
- Duration: One-Dimensional (1D)
- Pose Constraints:
- One-Dimensional (1D) frame index of each constraint
- Features to constrain may include Three-Dimensional (3D) joint positions, (3x3) joint rotation matrices, Two-Dimensional (2D) heading direction, and/or Two-Dimensional (2D) root position
Other Properties Related to Input: Maximum duration is 10 sec (300 frames at 30 frames per second).
Outputs
Output Type: Skeleton Motion: Root Translation and Joint Rotations
Output Formats:
- Root Translation: Matrix
- Joint Rotations: Matrix
Output Parameters:
- Root Translation: Two-Dimensional (
num_frames x 3)
- Joint Rotations: Four-Dimensional (
num_frames x 30 x 3 x 3)
Other Properties Related to Outupt:
- Motions are at 30 frames per second (30 fps)
Our AI models are designed and/or optimized to run on NVIDIA GPU-accelerated systems. By leveraging NVIDIA’s hardware (e.g. GPU cores) and software frameworks (e.g., CUDA libraries), the model achieves faster training and inference times compared to CPU-only solutions.
Software Integration:
Runtime Engines:
Supported Hardware Microarchitecture Compatibility:
- NVIDIA Ampere
- NVIDIA Blackwell
- NVIDIA Lovelace
Supported Operating Systems:
The integration of foundation and fine-tuned models into AI systems requires additional testing using use-case-specific data to ensure safe and effective deployment. Following the V-model methodology, iterative testing and validation at both unit and system levels are essential to mitigate risks, meet technical and functional requirements, and ensure compliance with safety and ethical standards before deployment.
Model Version
Kimodo-SOMA-RP-v1.1
Training and Testing Datasets:
Name: Proprietary Bones Rigplay Dataset
Data Modalities
- Text
- Human Motion Capture
Data Size:
- Less than 1 Billion tokens of text
- 700 hours of human motion capture
Data Collection Method
Automatic/Sensors
Labeling Method
Hybrid: Automatic/Sensors, Human
Properties: 700 hours of captured human body motions on the SOMA skeleton with corresponding text descriptions. Test splits from the
Kimodo Motion Generation Benchmark are held out, while the rest is used for training. Various augmentations were employed to expand text and motion variety.
Quantitative Evaluation
For results on the benchmark, pleaser refer to the
Kimodo documentation.
Inference:
Acceleration Engine: N/A
Test Hardware:
- GeForce RTX 3090
- GeForce RTX 4090
- GeForce RTX 5090
- NVIDIA A100
- NVIDIA L40S
- NVIDIA L4
- NVIDIA RTX 6000 Ada
- NVIDIA RTX A6000
Ethical Considerations:
NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. When downloaded or used in accordance with our terms of service, developers should work with their internal model team to ensure this model meets requirements for the relevant industry and use case and addresses unforeseen product misuse.
For more detailed information on ethical considerations for this model, please see the Bias, Explainability, Safety & Security, and Privacy Subcards below.
Please report model quality, risk, security vulnerabilities or NVIDIA AI Concerns
here.
Bias
| Field | Response |
|---|
| Participation considerations from adversely impacted groups protected classes in model design and testing: | Gender |
| Measures taken to mitigate against unwanted bias: | Our training data contains motion captured from a roughly equal number of male and female actors |
Explainability
| Field | Response |
|---|
| Intended Task/Domain: | Character Animation |
| Model Type: | Diffusion Transformer |
| Intended Users: | The model is intended for users with any level of animation experience to create 3D human motion data for their application. This may include demonstrations for humanoid robots, digital human motion for simulations and synthetic data, and animations for games and media. |
| Output: | 3D skeletal animation (root translation and joint rotations) |
| Describe how the model works: | Text input and pose constraints are processed and given to a transformer-based model that iteratively denoises a sequence of body poses. |
| Name the adversely impacted groups this has been tested to deliver comparable outcomes regardless of: | Gender |
| Technical Limitations & Mitigation: | Generated motions may include artifacts like foot skating where feet slide unnaturally when they should be in static contact with the ground. The motion does not always follow the given text prompt, and the model does not know how to perform certain types of actions (e.g., the model is best at locomotion, gestures, combat, dancing, and everyday activities). Each trained model currently outputs motion for a single character skeleton. The model is designed to output realistic human motions, so it cannot create cartoon motions or non-physically plausible motions. The model is not aware of objects in the scene around a character. |
| Verified to have met prescribed NVIDIA quality standards: | Yes |
| Performance Metrics: | Pose Constraint Accuracy (joint distance error), Motion Quality (foot-skating error, FID, latent similarity), Text-Following Accuracy (R-precision, latent similarity) |
| Potential Known Risks: | The model may output body motions that inadvertently reflect stereotypes related to age, gender, or physical characteristics. To mitigate this, prompts should describe actions in neutral, physical terms (e.g., “A person walks slowly with shuffled steps”) rather than relying on demographic adjectives. |
| Licensing: | This model is released under the NVIDIA Open Model License |
Privacy
| Field | Response |
|---|
| Generatable or reverse engineerable personal data? | No |
| Personal data used to create this model? | No |
| How often is dataset reviewed? | During dataset creation, model training, evaluation and before release |
| Was data from user interactions with the AI model (e.g. user input and prompts) used to train the model? | No |
| Is there provenance for all datasets used in training? | Yes |
| Does data labeling (annotation, metadata) comply with privacy laws? | Yes |
| Is data compliant with data subject requests for data correction or removal, if such a request was made? | Not Applicable |
| Applicable Privacy Policy | https://www.nvidia.com/en-us/about-nvidia/privacy-policy/ |
Safety
| Field | Response |
|---|
| Model Application Field(s): | Media & Entertainment, Industrial/Machinery and Robotics, Autonomous Vehicles |
| Describe the life critical impact (if present). | Not Applicable |
| Use Case Restrictions: | Abide by the NVIDIA Open Model License |
| Model and dataset restrictions: | The Principle of least privilege (PoLP) is applied limiting access for dataset generation and model development. Restrictions enforce dataset access during training, and dataset license constraints adhered to. |
