Abstract: In the era of Industry 4.0, robot motion planning faces unprecedented challenges in adapting those high-dimension dynamic working environments with rigorous real-time planning requirements. Traditional sampling-based planning algorithms can find solutions in high-dimensional spaces but often struggle with achieving the balance among computational efficiency, real-time adaptability, and solution optimality. To overcome these challenges and unlock the full potential of robotic automation in smart manufacturing, we propose bidirectional recurrent motion planning network (BRMPNet). As an imitation learning-based approach for robot motion planning, it leverages deep neural networks to learn the heuristics for approximate-optimal path planning. BRMPNet employs the refined PointNet++ network to incorporate raw point-cloud information from depth sensors and generates paths with a bidirectional strategy using long short-term memory (LSTM) network. It can also be integrated with traditional sampling-based planning algorithms, offering theoretical assurance of the probabilistic completeness for solutions. To validate the effectiveness of BRMPNet, we conduct a series of experiments, benchmarking its performance against the state-of-the-art motion planning algorithms. These experiments are specifically designed to simulate common operations encountered within generic robotic platforms in smart manufacturing such as mobile robots and multi-joint robotic arms. The results demonstrate BRMPNet’s superior performance on key metrics including solution quality and computational efficiency, suggesting the promising potential of learning-based planning in addressing complex motion planning challenges.

The full text can be downloaded at https://link.springer.com/article/10.1007/s40436-024-00529-6

Key words: Robot motion planning, Imitation learning, Deep neural network, Smart manufacturing, Adaptive and real-time planning