A method and system for controlling a native brain-like large model of a robot with a body

Abstract

The application discloses a kind of original brain-like big model control method of embodied intelligent robot, it is related to industrial robot control technical field.The method is converted into vertical knowledge representation by industrial vertical semantic modeling to industrial scene data;Task decomposition result and action strategy are generated by original brain-like big model in combination with knowledge, experience and constraint;Virtual deduction is carried out in digital twin space and risk and deviation index are output;After safety gate control verification, it is issued to embodied executor real execution;Deviation feedback is generated by virtual-real data alignment, and model online self-evolution update is driven;Memory unit is stored in experience base at the same time, and similar experience is retrieved and reused;And mature experience is abstracted as skill module to realize cross-scene migration.The application solves the problems of complex industrial scene tasks, unreliable strategies, virtual-real disconnection, difficult reuse of experience, and difficult self-evolution of models, and realizes safe and reliable, self-learning, and transferable industrial-level embodied intelligent control, suitable for assembly, machining, handling, inspection and other industrial scenes.

Images