Deep Learning-driven Real-time Whole Body Obstacle Avoidance For Multi-DoF Redundant Manipulator In Space

Abstract

Redundant multi-degree-of-freedom (DoF) manipulators are indispensable in space operations, offering unparalleled dexterity and adaptability in dynamic and resource-constrained environments. This study introduces a novel real-time whole-body obstacle avoidance (RWOA) framework, seamlessly integrating deformable dynamical systems with deep learning algorithms to address the challenges posed by microgravity, computational constraints, and dynamically moving obstacles. The proposed framework leverages a combined modulation matrix to adaptively deform the system's dynamics, enabling collision-free trajectories across the manipulator's entire structure, while ensuring precise trajectory tracking for mission-critical tasks. Deep learning is incorporated to predict obstacle motion patterns, facilitating anticipatory and adaptive control strategies that enhance operational efficacy in high-dimensional and unpredictable environments. The efficiency of the proposed method is further demonstrated through the development of a lightweight convolutional neural network architecture capable of achieving real-time obstacle trajectory prediction with minimal computational overhead. Extensive simulations and experimental validations in zero-gravity conditions reveal significant advancements in trajectory optimization, obstacle avoidance, and overall manipulator performance. This framework offers a transformative approach to robotic manipulation in extraterrestrial environments, establishing a robust foundation for future advancements in space robotics and adaptive control.