nvidia/Kimodo-SOMA-RP-v1.2

Kimodo: Controllable Kinematic Motion Diffusion at Scale

[Paper](https://research.nvidia.com/labs/sil/projects/kimodo/assets/kimodo_tech_report.pdf), [Project Page](https://research.nvidia.com/labs/sil/projects/kimodo/)

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-v1.2, an updated Kimodo-SOMA-RP model. This model is ready for commercial use.

Changes in v1.2:

Kimodo-SOMA-RP-v1.2 is trained on corrected training splits: in v1.1, some BONES-SEED test sequences had leaked into the training set, so its numbers on the Motion Generation Benchmark are not fully comparable to models trained on the recommended splits. v1.2 removes the leaked sequences, making it the model to compare against on the benchmark. Compared to v1.1 it reduces foot skating and matches constraint accuracy; its text-following scores (R-precision/FID) are slightly lower, partly reflecting the removed leakage. See the benchmark results page in the codebase for the full tables.

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 [10/15/2026] via link

HuggingFace [10/15/2026] via link

References:

• Technical report: Kimodo: Scaling Controllable Human Motion Generation
• Webpage: link

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:

• PyTorch

Supported Hardware Microarchitecture Compatibility:

• NVIDIA Ampere
• NVIDIA Blackwell
• NVIDIA Lovelace

Supported Operating Systems:

• Linux
• Windows

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-v1.2

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. Split into 90%/10% train/test splits. Various augmentations were employed to expand text and motion variety.

Quantitative Evaluation

For test set evaluation, please refer to the benchmark results in the repo.

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,...