Adaptive Wall-Following Control for Unmanned Ground Vehicles Using Spiking Neural Networks

TL;DR

The paper addresses robust wall-following for UGVs under modeling uncertainties and disturbances by integrating real-time B-spline wall fitting, a discretized LQR baseline, a feedforward point-matching module, and an adaptive SNN controller trained online via the Prescribed Error Sensitivity rule. This hybrid control scheme enables accurate trajectory tracking and wall contour following even with actuator faults and state estimation errors, outperforming a traditional LQR in simulations. The key contributions include a practical wall-fitting reference generator, a real-time MPF-based feedforward term to accelerate convergence, and an online-adaptive SNN that continuously updates weights to minimize position and heading errors. The results demonstrate improved tracking accuracy and convergence speed, suggesting practical significance for UGVs operating in uncertain, cluttered, or unknown environments, potentially leveraging neuromorphic hardware for efficiency.

Abstract

Unmanned ground vehicles operating in complex environments must adaptively adjust to modeling uncertainties and external disturbances to perform tasks such as wall following and obstacle avoidance. This paper introduces an adaptive control approach based on spiking neural networks for wall fitting and tracking, which learns and adapts to unforeseen disturbances. We propose real-time wall-fitting algorithms to model unknown wall shapes and generate corresponding trajectories for the vehicle to follow. A discretized linear quadratic regulator is developed to provide a baseline control signal based on an ideal vehicle model. Point matching algorithms then identify the nearest matching point on the trajectory to generate feedforward control inputs. Finally, an adaptive spiking neural network controller, which adjusts its connection weights online based on error signals, is integrated with the aforementioned control algorithms. Numerical simulations demonstrate that this adaptive control framework outperforms the traditional linear quadratic regulator in tracking complex trajectories and following irregular walls, even in the presence of partial actuator failures and state estimation errors.

Figures (9)

• Figure 1: Framework of the adaptive SNN control design.
• Figure 2: Single-layer SNN architecture.
• Figure 3: Leaky Integrate-and-Fire (LIF) neuron model.
• Figure 4: Case Study A: straight-line tracking with actuator failure. The adaptive SNN controller converges more quickly to the desired trajectory and covers a greater distance compared to the benchmark LQR controller.
• Figure 5: Case Study A: the error signals of the adaptive SNN converge to zero at $t=17$s, whereas the LQR fails to achieve convergence by the end of the simulation.