Abstract
Accurate orientation estimation is essential for mobile robots, yet common sensing methods such as
encoders and IMUs accumulate drift over time. As a low-cost alternative that emulates the behaviour
of an optical rotary encoder, this work employs the robot’s built-in infrared (IR) reflectance sensors
together with a custom circular greyscale gradient to infer orientation directly from reflectance
patterns. Two approaches are evaluated: a mathematical curve-fitting method and a neural network-
based model. The curve-fitting approach filters IR data using cascaded Chebyshev Type II IIR biquads
and fits a sinusoidal model via the Levenberg-Marquardt algorithm, while the neural network uses a
compact multilayer perceptron to perform end-to-end angle prediction. Experimental results show
that both methods achieve comparable average accuracy, with the neural network offering improved
performance under darker conditions, whereas the curve-fitting method remains more efficient in
memory and computation.