Learning-based path planning for smart robotic industrial applications

This thesis investigates learning-based approaches to improve the performance of path planning algorithms in structured and quasi-static environments. The objective is to integrate learning mechanisms with classical sampling-based planning techniques in order to reduce planning time while preserving reliable obstacle avoidance. Two complementary planning paradigms are analyzed. The first approach is based on the Zonal RL-RRT framework, which combines reinforcement learning with Rapidly-exploring Random Trees (RRT). The workspace is partitioned into spatial zones through a kd-tree decomposition, and a high-level policy learned via Q-learning selects a sequence of zones connecting the start and goal regions. Local RRT planners are then used to generate collision-free trajectory segments between consecutive zones. The second approach investigates trajectory generation using Dynamic Movement Primitives (DMPs), where robot motions are modeled as stable dynamical systems whose parameters can be learned from data. In this framework, neural networks are employed to predict trajectory parameters, allowing fast generation of smooth motions adapted to different obstacle configurations. The proposed approaches are evaluated in simulation through robotic manipulation tasks inspired by industrial palletization scenarios. Experiments are conducted in a physics-based environment using the MuJoCo engine and a UR10e robotic manipulator. Performance is assessed in terms of computation time, trajectory length, and motion characteristics. The results highlight the strengths and limitations of both paradigms. Reinforcement learning guided sampling improves exploration efficiency in cluttered environments, whereas DMP-based trajectory generation enables faster inference and smoother motions. The analysis provides insights into the trade-offs between sampling-based planning and learning-based motion generation for industrial robotic applications.