Train a diffusion model to generate humanoid motion from text descriptions or audio.
Export a trained motion model to ONNX for fast GPU inference.
Run a combined generation and tracking pipeline to produce robot control signals.
Deploy the generative planner and tracking system to a Unitree G1 humanoid robot.
| tsinghua-mars-lab/omg | eadmin2/jarvis_ai | greatvishal27-rc/ai-resume_analyzer | |
|---|---|---|---|
| Stars | 56 | 56 | 56 |
| Language | Python | Python | Python |
| Setup difficulty | hard | hard | moderate |
| Complexity | 5/5 | 4/5 | 2/5 |
| Audience | researcher | developer | vibe coder |
Figures from each repo's GitHub metadata at analysis time.
Requires multiple CUDA GPUs for training and a Unitree G1 with Nvidia Orin for real-robot deployment, some checkpoints are not yet released.
OMG (Omni-Modal Motion Generation for Generalist Humanoid Control) is a research project from Tsinghua University's MARS Lab that trains AI models to generate realistic movements for humanoid robots. The work is tied to an academic paper and focuses on getting a robot to move in response to different kinds of instructions, such as text commands or audio cues. The core approach uses a diffusion model, a type of AI that learns to produce outputs by progressively refining random noise into something structured. Here, the model learns to produce sequences of body joint positions that form coherent motions. Models of several sizes are provided, ranging from 50 million to 1 billion parameters, and training can be distributed across multiple GPUs. The workflow has several stages. You download the training data and pretrained model weights (some of which are not yet released at the time of this writing), preprocess the data into fixed chunks for efficiency, compute normalization statistics, and then train the diffusion model. After training, you export the model to ONNX format for fast inference using TensorRT on a CUDA-capable GPU. A separate motion-tracking component called HoloMotion converts the generated joint positions into actual robot control signals. For running the system, the repository supports several pipeline modes: generating motion only, tracking only, or a combined online or offline pipeline that chains generation and tracking together. The generation step accepts conditions such as text descriptions and a seed motion clip. Real-robot deployment targets the Unitree G1 humanoid running on an Nvidia Orin compute module. A GPU workstation runs the generative planner while the robot's onboard system handles motion tracking and actuation. The project is MIT licensed.
A research project that trains diffusion models to generate humanoid robot motion from text or audio commands.
Mainly Python. The stack also includes Python, CUDA, TensorRT.
Use freely for any purpose, including commercial use, as long as you keep the copyright notice.
Setup difficulty is rated hard, with roughly 1day+ to a first successful run.
Mainly researcher.
This repo across BitVibe Labs
Verify against the repo before relying on details.