Learning-based Approximate Model Predictive Control for an Impact Wrench Tool

TL;DR

<3-5 sentence high-level summary> This work tackles the challenge of real-time torque control for impact wrenches under tight embedded-resource constraints. It introduces a learning-augmented MPC that augments a nominal model with Gaussian-process residuals and then uses a neural network to approximate the MPC policy for microsecond-level, real-time inference on embedded hardware, with EKF-based state estimation. The approach yields high constraint satisfaction and improved tracking compared with a baseline PID controller, achieving 97.99% closed-loop success over thousands of trials and a 490× computational speedup over exact MPC. Hardware experiments validate real-time deployment and constraint adherence, making the framework practically impactful for safety-critical, high-frequency mechatronic control.

Abstract

Learning-based model predictive control has emerged as a powerful approach for handling complex dynamics in mechatronic systems, enabling data-driven performance improvements while respecting safety constraints. However, when computational resources are severely limited, as in battery-powered tools with embedded processors, existing approaches struggle to meet real-time requirements. In this paper, we address the problem of real-time torque control for impact wrenches, where high-frequency control updates are necessary to accurately track the fast transients occurring during periodic impact events, while maintaining high-performance safety-critical control that mitigates harmful vibrations and component wear. The key novelty of the approach is that we combine data-driven model augmentation through Gaussian process regression with neural network approximation of the resulting control policy. This insight allows us to deploy predictive control on resource-constrained embedded platforms while maintaining both constraint satisfaction and microsecond-level inference times. The proposed framework is evaluated through numerical simulations and hardware experiments on a custom impact wrench testbed. The results show that our approach successfully achieves real-time control suitable for high-frequency operation while maintaining constraint satisfaction and improving tracking accuracy compared to baseline PID control.

Figures (4)

• Figure 1: The internal mechanism of the impact wrench. The motor torque drives the spindle rotation, which transmits force to the hammer through the ball-and-cam mechanism. The spring stores and releases energy during the impact cycle.
• Figure 2: Impact wrench operational modes. Red line: hammer's lowest position; blue region: anvil. (a) Desirable impact configuration with direct hammer-to-anvil foot contact. (b) Failure mode exhibiting premature axial contact at multiple anvil surfaces, resulting in reduced torque output, increased component wear, and excessive vibrations.
• Figure 3: Simulation comparison of the Neural Network (NN), Model Predictive Control (MPC), and a baseline speed controller. (a) Spring Angle at Impact: The NN (blue line) and MPC (green line) controllers track the 0.2 rad reference (black dashed). (b) Max Hammer X Position: Both the NN and MPC controllers respect the 0.012 m constraint (black solid), while the speed controller violates it.
• Figure 4: Hardware experimental results showing constraint satisfaction during controller transition. Red shaded region: speed Controller (4.5--5.0s); Blue shaded region: NN Controller (5.0--6.0s). (a) Spring angle tracking demonstrates improved reference following under NN. (b) Maximum hammer position remains within safety constraints throughout both control phases.