Sim-to-real robotic grinding control via physics-enhanced learning and predictive optimization for unknown surface features

This is the authors' abstract. We don't add key points for this paper.

Unknown and unstructured wall features often lead to inaccurate recognition and insufficient end-effector contact during robotic grinding, making it difficult to maximize grinding efficiency without compromising surface flatness. Additionally, reproducing diverse wall features and fairly comparing control strategies remain challenging in real-world experiments. To address these challenges, this paper proposes a Dynamic Optimal Grinding Control Framework (DOGCF) to achieve highly efficient and robust robotic wall grinding. First, we establish a classification-driven mapping framework to extract finite-feature primitives that represent diverse wall topographies. Then, a Physics-Enhanced Long Short-Term Memory Network (PE-LSTM) is proposed to facilitate the recognition of wall features and the prediction of the end-effector’s pose using multidimensional force/position data. It is trained on about 96% simulated data and 4% real-world experimental data, totaling approximately 19,200 simulated trajectories and 800 real trajectories. Additionally, we integrate the PE-LSTM outputs and a feature-removal model into a Model Predictive Control (MPC) framework, resulting in a locally optimal sequence of grinding actions. Comparative experiments show that our method achieves an average removal compliance rate of 90.7% and a mean overall wall flatness of 1.88 mm, which represent improvements of 5.43% and 7.15%, respectively, compared to the best of three representative methods. In addition, the grinding efficiency increased by 17.4%. The PE-LSTM and DOGCF proposed in this paper have potential for robotic grinding of unknown large surfaces when integrated with workspace-extension mechanisms.