A robot path tracking method based on manual guidance and path reinforcement learning

Controlling the movement of an industrial robot along specific edges of a workpiece in a complex environment, where multiple paths intersect, is crucial for tasks such as welding and gluing. Traditional robot teaching methods restrict robots to fixed task environments using pre-programmed motion planning schemes. Although vision-guided robotic path-tracking systems can automatically extract paths, the presence of multiple intersections complicates autonomous path determination and tracking using conventional vision-based algorithms. To address this challenge, this study proposed a robot path-tracking approach that integrates manual guidance with path reinforcement learning. This strategy leverages both visual- and human-guided information to learn complex manipulation skills that require precise positional constraints and continuous motion, such as welding or gluing, in environments with intersecting paths. A user-friendly robot path teaching framework was designed, allowing operators to select key positions on the robot manipulator's motion path (2D guide pixel points) from color images using a mouse to generate guide images. However, these interactively selected 2D guide pixel points may introduce biases relative to the ideal robot path (i.e., the edge of the workpiece that needs to be tracked). To mitigate this, a path reinforcement learning technique was proposed that uses the edge image of the workpiece along with manual guidance to determine the necessary actions (2D pixel tracking path points) for tracking specific edges in complex environments. This process is constrained by guide images and an invalid action mask matrix. An invalid action mask matrix, calculated from the guide points, prevents the exploration of suboptimal trajectories during path reinforcement learning. The robot's 6- degrees of freedom (DOF) path was then derived from the 2D pixel-tracking path points and depth images. Finally, the accuracy of 2D pixel path tracking was tested in a virtual environment, yielding an average error of 0.363 pixels and a standard deviation of 0.594 pixels. The effectiveness of the proposed path-tracking approach in scenarios with multiple intersecting paths was verified in a physical environment.