A decision control method, system, device and medium of a humanoid robot

Abstract

This invention discloses a decision-making and control method, system, device, and medium for humanoid robots. The unified policy model in this method is trained through the following steps: acquiring several full-dimensional training features of the humanoid robot; training a single-dimensional policy model based on each full-dimensional training feature to obtain several trained single-dimensional policy models, and original trajectory information corresponding to each trained single-dimensional policy model; the original trajectory information includes state behavior data, action behavior data, and reward behavior data; unifying the reward across scenarios for all original trajectory information to obtain a unified trajectory dataset; and training an initialized unified policy model based on the unified trajectory dataset to obtain a trained unified policy model. The model provided by this method can effectively improve the scenario adaptability of humanoid robot decision-making and control. This invention relates to the field of artificial intelligence technology.

Images