On June 24, Chengdu Humanoid Robot Innovation Center released R-DDPRM, China’s first humanoid robot task generation model based on visual diffusion architecture, which allows robots to “silkily” handle various complex tasks without hesitation, which is a major core technological breakthrough in the evolution history of humanoid robots.

DRG3-1008B-230B “Chengdu Humanoid Robot Innovation Center is committed to building humanoid robots with ‘brains’ of international leading level.” Zhang Ruirui, head of Chengdu Humanoid Robot Innovation Center Co., LTD., said: “Some of the well-known models are vertical models based on diffusion architecture, and they have strong understanding, reasoning and generation capabilities without exception. The application of diffusion architecture to the research and development of humanoid robot core technology is the latest and most cutting-edge field of global artificial intelligence, which will reconstruct the underlying architecture of humanoid robots, release the decision-making skills of humanoid robots, and make it reach an unprecedented height in terms of manipulation and generalization. At present, only Tesla, MIT + Columbia University joint team and other top international humanoid robot teams in the world have released relevant phased results.”

DRG3-1008B-230B

“Current industry task generation models for humanoid robots are mainly based on VAE, AC Transformer, LSTM and other architectures, but humanoid robots under such architectures usually consider a single constraint type when planning actions, such as object holding stability or obstacle avoidance path planning. This requires the development and execution of a series of separate vertical models (parallel linear) for each subdivision, and implementation problems require local adjustment of trial and error to slowly advance the task process, resulting in low reasoning efficiency, slow execution, and large consumption of computing power.” Zhang Ruirui introduced that the R-DDPRM model launched by Chengdu Humanoid Robot Innovation Center can allow humanoid robots to generalize across multiple constraints, creating a real humanoid robot “brain” – able to “silkily” handle a variety of complex tasks without hesitation.

DRG3-1008B-230B The model has the ability of global planning thinking to determine the best holding position of the object, the best route of moving the object and the best energy consumption. Moreover, it can quickly make decision corrections and adjust the execution mode when the task process is disrupted by humans, and continue to complete the task, and the generalization ability of the model allows the humanoid robot to smoothly complete a variety of complex tasks from cleaning rooms to complex surgical operations.

According to reports, R-DDPRM task generative model has absolute advantages such as more stability, greatly improved task execution success rate, more agility in decision-making, geometrically reduced computing power dependence (theoretically, the middle consumer graphics card can complete most tasks of humanoid robots on the end), and stronger ability to handle complex scene tasks. It will greatly improve the task execution ability and efficiency of humanoid robots, and reduce the cost of computing power and energy consumption.

“The release of R-DDPRM will realize the qualitative leap of humanoid robots from parallel linear reasoning to agile global generalization reasoning in task generation, which is a breakthrough in the field of humanoid robots and will redefine the underlying logic of humanoid robot reasoning and task planning.” Zhang Ruirui said.