Trajectory smoothing method using reinforcement learning for computer numerical control machine tools

Abstract

Abstract Tool-path codes output by computer-aided manufacturing software for high-speed machining are composed of discontinuous G01 line segments. The discontinuity of these tool movements causes computer numerical control (CNC) inefficiency. To achieve high-speed continuous motion, corner smoothing algorithms based on pre-planning methods are widely used. However, it is difficult to optimize smoothing trajectories in real-time systems. To obtain smooth trajectories efficiently, this paper proposes a neural network-based direct trajectory smoothing method. An intelligent neural network agent outputs servo commands directly based on the current tool path and running state in every cycle. To achieve direct control, motion feature and reward models were built, and reinforcement learning was used to train the neural network parameters without additional experimental data. The proposed method provides higher cutting efficiency than the local and global smoothing algorithms. Given its simple structure and low computational demands, it can easily be applied to real-time CNC systems.