DA-MRL: Dual-Arm Meta-Reinforcement Learning for Fast Adaptation in Dynamic Obstacle Avoidance

Dynamic obstacle avoidance in a shared dual-arm workspace requires coordinated motion and rapid adaptation to changing task conditions. This paper presents dual-arm meta-reinforcement learning (DA-MRL), a centralized framework that learns a task-generalized actor--critic initialization for adaptation to unseen target, obstacle, sensing, and dynamic conditions using limited support trajectories. MuJoCo experiments evaluate held-out in-distribution (ID) and out-of-distribution (OOD) tasks in both three-degree-of-freedom (3-DOF) and six-degree-of-freedom (6-DOF) systems. After ten adaptation updates, DA-MRL increases OOD collision-free success from 36.0% to 68.4% in 3-DOF and from 28.6% to 55.1% in 6-DOF, while reducing the corresponding collision rates. The adapted 6-DOF policy is then frozen and transferred to a physical dual-UR5 platform with red-green-blue and depth (RGB-D) perception. None of the six continuous 200-s deployment records, each containing 20,000 logged samples, encountered the defined collision-induced termination condition. These results demonstrate rapid policy adaptation in simulation and coordinated sim-to-real execution under the evaluated dynamic obstacle conditions.