The Zhongke Semiconductor team has launched the first GaN based programmable embodied robot power system chip, which adopts SIP packaging technology and is equipped with a hardware acceleration engine, high-speed interface, PWM signal array programmable unit, edge image processing, and high-speed interface for various sensors and biological information acquisition. Mainly used in the field of multi joint embodied robots and intelligent equipment.

The artificial intelligence power system chip released by Zhongke Semiconductor this time controls 100 simulated muscle and servo motor systems through the array PWM current signal output by the chip's built-in "edge physics model" (referring to the constraints of deep learning and physical laws, energy conservation, sensors and vision added to the neural network through data-driven methods) based on inference large models (such as ChatGPT, open AI, DeepSeek generated 3D virtual models and pose coordinates) to complete complex atomic operations, with a single pose motion of 32 degrees of freedom. By using the "edge physics model" to output PWM signals to control physical quantity signals, the robot can have complex limb movements and physical quality constraints similar to humans.

The company's research team has been deeply engaged in this field for many years, and this chip technology has obtained multiple invention patents and won the second and third prizes in the ICCV International Artificial Intelligence Competition in the past two years. In 2024, it won the international artificial intelligence visual tactile CVPR championship.

In recent years, artificial intelligence has developed rapidly, especially since 2018, with the rise of generative artificial intelligence driving model training and the construction of a large number of algorithm centers. In less than two years, the "model training phase" has been completed. Within less than three years, language inference models such as ChatGPT, openAI, and DeepSeek were successively launched. The era of artificial intelligence has undergone a rapid evolution from the "training stage" to the "inference stage" of models. With the large-scale commercial use of inference models in China by 2025, the arrival of the "artificial intelligence application stage" has also begun. Compared to hardware, AI chips are quietly changing in the "model training stage" around the rapid evolution of inference big models. From foreign chip manufacturers such as NVIDIA's GPUs to Broadcom's recent launch of the latest AI ASIC (Application Specific Integrated Circuit) chips to meet the needs of different fields of artificial intelligence applications, coupled with China's strong promotion of artificial intelligence+low altitude economy construction, there is an opportunity for the development of AI ASIC dedicated chips. The company will continue to deepen its research and development of chips and "AI physical models" in this field.