Neural Autonomous Dynamical System for Robot Learning and Obstacle Avoidance

Recent advancements in learning from demonstration algorithms have significantly enhanced human–robot skill transfer, enabling robots to directly acquire skills through human demonstrations, thereby accelerating the process of task learning and execution. Learning from demonstration based on autonomous dynamic systems has been widely applied and proven to be an efficient and reliable approach, particularly due to its time independence. However, balancing the complexity, accuracy, and generalization of dynamic systems remains a challenging problem. This article proposes a novel method based on neural networks for the efficient reproduction and generalization of human demonstrations. This method constructs a modulation matrix through neural networks to automatically learn the trajectory features and shapes of human demonstrations, and incorporate a neural energy function to ensure the stability of the system. The integration of the energy function and modulation matrix not only guarantees stability but also enhances generalization performance. For tasks in complex environments with obstacles, we further introduce obstacle information and decision functions into the modulation strategy, enabling the robot to learn human-demonstrated obstacle avoidance behaviors and achieve autonomous obstacle avoidance and task completion. The effectiveness and superiority of the proposed method are validated through simulation experiments on a 2-D dataset and robot obstacle avoidance tasks.