Recurrent neural network-based dynamic obstacle avoidance for dual-arm robot with joint discomfort optimization

Abstract Humanoid dual-arm robots face significant challenges in planning safe and humanoid motions during collaborative tasks due to overlapping workspaces and kinematic redundancy. This paper proposes a real-time joint adjustment planning method that optimizes dual-arm motion by defining a “joint discomfort” metric, which quantifies the deviation of joint angles from their comfort positions, thereby enhancing the naturalness and human-likeness of the motion. Leveraging the parallel computing capability of recurrent neural networks (RNNs) and the redundancy resolution of dual arms, our method dynamically adjusts joint configurations to minimize discomfort while integrating trajectory tracking and obstacle avoidance constraints. Predefined tasks and constraints are formulated as a Quadratic Programming (QP) problem, efficiently solved using an RNN-based approach. Numerical simulations and physical experiments on the Ginger robot – a dual-arm system with 7-degree-of-freedom (DOF) manipulators – validate the efficacy of the proposed planning method in three key aspects: (1) optimizing joint space utilization to enhance workspace flexibility, (2) adaptively regulating joint motions, and (3) achieving sub-millimeter tracking accuracy (position error ¡0.15 mm) under dynamically constrained scenarios.