On May 29, this morning, the National and Local Co-constructed Humanoid Robot Innovation Center and the Future Information Innovation Institute of Fudan University officially released the world's first generative humanoid robot motion model - "MindLoong" (MindLoongGPT).
The MindLoong model takes "natural language drive" as its core, and builds a complete closed loop from multimodal input to high-fidelity motion generation, subverting the traditional motion control paradigm.
In other words, users do not need to learn professional terms or operate complex software. They only need to issue instructions like talking to humans, such as "handing a cup of coffee in an elegant posture", "waving" or uploading a reference video. The MindLoong model can automatically parse the semantics and generate corresponding coherent actions, and support real-time adjustment of details by adding instructions.
This ability is subversive in the fields of education and service robots. It is reported that the MindLoong model has four core technical advantages:
First, multimodal interaction and low-threshold drive. Users only need to drive with natural language, and the model can "listen to all actions".
Second, the simulation degree and timing consistency are double breakthroughs. By considering human movements as "timing language", the generated running, dancing and other movements are not only smooth and natural, but also retain the inertial rhythm of the real human body.
Third, both structure generation and detail optimization are given equal importance. The model adopts a phased generation strategy, achieving millimeter-level precision control from global posture to local joint angles.
Fourth, lightweight design enables a wide range of scenarios. While ensuring performance, the volume of the MindLoongGPT model is compressed to 1/3 of the same type of products, and can run in real time on embedded platforms such as educational robots and smart wearable devices, paving the way for industrial implementation.