Momentum-Incorporated Discrete Neural Network for Learning and Control of Redundant Robots

Hierarchical control is widely used in redundant robotic control systems to decompose complex tasks into manageable layers. However, conventional hierarchical control methods typically possess two drawbacks. First, they are constrained by the assumption that the total task dimension is less than or equal to the number of degrees of freedom (DOFs) of a redundant robot. Second, they rely on precise physical parameters, and the control accuracy degrades under parametric uncertainties. To overcome these drawbacks, this article proposes a momentum-incorporated discrete neural network (MIDNN) for multitask hierarchical control in an overdetermined system with uncertain physical parameters. The MIDNN incorporates two subnetworks based on a momentum-incorporated discrete (MD) method: a multitask hierarchical control network (MHCN) and a physical parameter learning network (PPLN). The MHCN enables priority-based multitask hierarchical control in scenarios where the total task dimension exceeds the number of DOFs. Meanwhile, the PPLN is designed to learn uncertain physical parameters, thus improving network adaptivity. The MD method accelerates the convergence of the network. Theoretical analyses confirm the convergence of the MIDNN, while both simulations and physical experiments validate the superiority in hierarchical control, physical parameter learning, and convergence acceleration of the proposed MIDNN. The accompanying video can be accessed at https://youtu.be/fUsva4xJWhM.