Adaptive process parameters decision-making in robotic grinding based on meta-reinforcement learning

Abstract

In robotic grinding, the variability of workpiece characteristics, uneven machining allowances and nonlinear tool wear collectively pose challenges for consistent material removal. To address these dynamic grinding conditions and achieve high-accuracy material removal, this paper presents an adaptive decision-making model for grinding process parameters based on the meta-reinforcement learning. The proposed approach accurately adjusts grinding process parameters under a wide range of coating characteristics, multiple grinding tool types and progressive tool wear, with few-shot training samples. First, we develop an enhanced proximal policy optimization algorithm with better experience (PPO_BE) to optimize process parameters for a specific grinding task, improving material removal accuracy. Subsequently, building on the PPO_BE framework, we integrate model-agnostic meta-learning (MAML) to form MAML-PPO_BE algorithm, enabling fast adaptation across heterogeneous grinding tasks while preserving high accuracy. Comprehensive experiments on 16 distinct grinding tasks demonstrate a 51.4 %–68.9 % improvement in material removal deviation relative to the MAML, PPO_BE, SAC and FLC algorithms, respectively. This paper presents an adaptive parameters decision-making method with high accuracy in changing and complex grinding process.