Decoupling Sampling and Planning Frameworks for Redundant Manipulators Motion Planning Under End-Effector and Performance Constraints

We address the underexplored problem of motion planning for redundant manipulators under end-effector and performance constraints through motion planning on the intersection of constraint manifolds and regions (MPICMR). First, we propose a sampler framework called conditional generative adversarial network (CGAN) with an intersection-enhanced module (IE Module), referred to as CGAN-IE, to achieve efficient intersection-enhanced sampling. The IE Module with a dynamic attenuation strategy focuses the generator on constraint-satisfying samples at the intersections, improving sampling accuracy, while a diversity loss term is designed to alleviate the mode collapse of the basic CGAN, enabling the CGAN-IE to accurately capture the multimodal distribution of configurations at the intersections. CGAN-IE exhibits good generalization capability, generating new configurations for similar unseen planning scenarios without retraining. Second, based on CGAN-IE, we propose an MPICMR framework that decouples sampling and planning, allowing planners with different characteristics to be selected according to requirements, such as RRT*, EIT*, and BIT*, to achieve asymptotic optimality. Third, simulation experiments demonstrate the superiority of CGAN-IE: it maintains a high sampling accuracy at the sampling level, it exhibits stable generalization to similar unseen planning scenarios, and the planner performance and planning quality at the planning level are improved. Finally, physical experiments validate the generalization capability of the CGAN-IE sampler framework and the effectiveness of the MPICMR framework.