Al-Serri, Sari, Chalak Qazani, Mohamadreza, Mohamedi, Shady, Lim, Chee Peng, Nahavandi, Saeid, and Asadi, Houshyar (2025) Deep Q-Network for Optimising the Weights of Model Predictive Control-based Motion Cueing Algorithm. In: Proceedings of the IEEE International Conference on Systems Man and Cybernetics. pp. 3204-3211. From: SMC 2025: IEEE International Conference on Systems, Man, and Cybernetics, 5-8 October 2025, Vienna, Austria.

PDF (Published Version) - Published Version Restricted to Repository staff only

Abstract

Motion cueing algorithm aims to replicate realistic motion sensations for drivers while adhering to the physical limitations of the simulation platform. Model Predictive Control has been extensively employed within the domain of motion cueing algorithms for vehicle and flight simulators due to its ability to handle system constraints and optimise motion fidelity. However, traditional Model Predictive Control-based motion cueing algorithm rely on manually tuned cost function weights, which can be suboptimal and difficult to determine for different operating conditions. This suboptimal tuning can cause discrepancies between the visual input perceived by the simulator driver and the motion cues processed by their vestibular system, potentially resulting in motion sensation errors and increased risk of motion sickness. In this paper, we propose a reinforcement learning weight optimisation approach for the model predictive control-based motion cueing algorithm, leveraging Deep Q-Networks to determine an optimal set of cost function weights through training in a simulated environment. The optimised weights aim to minimise the cost function, thereby maximising the reward function. Simulation results indicate that the proposed method outperforms the traditional approach, leading to a reduction in motion sensation errors between the simulator and real vehicle driver and improving platform utilisation. The reinforcement learning-based control achieves better correlation between the reference and simulated signals for both sensed specific force and angular velocity, enhancing overall motion fidelity. It also reduces the root mean square error for sensed specific force, ensuring more accurate replication of target motion cues. Additionally, the method enables broader use of the simulator's linear displacement range, confirming the effectiveness of reinforcement learning in tuning control parameters for superior simulation performance.

Actions (Repository Staff Only)

Item Control Page