The invention discloses a dexterous hand operation strategy learning and cross-domain migration
algorithm for a rich contact task, and belongs to the field of
robot control and
artificial intelligence. The method comprises the following steps: firstly, constructing a generative
encoder network, and compressing a high-dimensional dexterous hand state to a probability
potential space as a unified strategy representation; then, in the space, human demonstration
semantics are used as guidance, an implicit dynamical model, a value function and a strategy network are jointly learned, the strategy network is used for preheating an optimization controller based on
model prediction path integration, and zero sample generalization of a target domain is achieved; and finally, through a progressive updating mechanism, a small amount of data of the target domain is utilized to finely adjust the
encoder, the domain difference is quickly captured, and adaptive migration of the strategy is realized. According to the method, the problem that the dexterous hand migrates from
simulation to reality due to dynamic domain difference in a rich contact task is effectively solved, and the method has the advantages of high zero sample generalization ability and high adaptive migration speed.