Mobile robots performing tasks in unknown environments need to traverse a variety of complex terrains,
and they must be able to reliably and quickly identify and characterize these terrains to avoid getting
into potentially challenging or catastrophic circumstances. However, currently, methods based on
different sensors, such as vision, lidar, audio, inertial measurement units (IMU), and tactile sensors
generally require huge computing resources for data processing, online training, and recognition, and
the performance tends to degrade once there are outlier data that do not exist in the prior training
dataset. To solve this problem, inspired by the animals, such as rats and seals, just relying on whiskers
to perceive information about their surroundings and survive in dark and narrow environments, we
explore the physical whisker-based reservoir computing for quick and cost-efficient mobile robots’
environment perception and navigation.
This thesis has creatively solved a machine-learning terrain classification problem by using a
tapered ‘electro-mechanical’ whisker-based reservoir computing system for the first time in the world.
The first analysis comprises designing a straight whisker sensor with one Hall sensor in the base
and evaluating its nonlinear interaction dynamics which will have one dominant frequency at the
vertical perturbation frequency of the sensor base. This nonlinear dynamic feature is used in a deep
multi-layer perceptron neural network to classify terrains. We achieved an 85.6% prediction success
rate for seven flat terrain surfaces with different textures at 0.2m/s. Therefore, a tapered whisker-based
reservoir computing (TWRC) system using a tapered whisker sensor with three Hall sensors along the
axis is proposed. The results demonstrated that such compliant tapered mechanical whisker systems
could achieve real-time perception using nonlinear vibration dynamics and providing morphological
computation power to achieve frequency separation in the time domain simultaneously. Then, by
running a numerical analysis and experiments, it was found that external terrain stimuli of different
roughness and hardness will produce unique whisker reservoir features and different whisker axis
locations and motion velocities provide variable dynamical response information. We achieved a
prediction success rate of 94.3% for six terrain surface classification experiments and 88.7% for
roughness estimation of the unknown terrain surface at a steady speed of 0.2m/s.Moreover, a tapered whisker-based semi-supervised reservoir computing (TWSSRC) system
is proposed to reveal that the whiskered robot can learn from prior physical experiences through
cost-efficient self-supervised reservoir computing to achieve auto-labelling of new terrain, terrain
classification, and terrain roughness estimation. The experimental results show that this novel approach is capable of successfully adapting well to unknown terrains based on the tapered whisker reservoir
outputs and detecting new terrains with high accuracy, achieving 84% accuracy over six terrains
with carpet representing the new terrain class. Depending on the computational superiority of the
TWSSRC system, an active improved-TWMC algorithm including an overlapping-window-based
decision module is designed, which could achieve active rapid object classification, even for the
extremely similar sandpaper including external disturbances. Finally, a real-time terrain identificationbased
navigation method is proposed using an onboard tapered whisker-based reservoir computing
(TWRC) system rather than an external computer. The results experimentally demonstrate that our
proposed algorithm could cost-efficiently achieve highly accurate real-time terrain classification results,
and analyzed and demonstrated experimentally how the mobile robot can be controlled by speed to
elicit unique frequency domain responses in a whisker sensor to help surface identification.
The research presented in this thesis demonstrates the exceptional performance of the reservoir
computing system based on tapered-shaped whisker sensors in environmental perception, particularly
in terrain recognition. This method provides a computationally efficient way to achieve high-precision
terrain recognition and classification, as well as semi-supervised labelling of unknown terrain, texture
information recognition, and roughness estimation. Additionally, this research sheds light on the importance
of natural body dynamics, especially the tapered-shaped whiskers, in addressing challenging
signal processing problems in robotics. It can also bridge the gap between high information processing
performance and low energy dissipation of robots’ onboard hardware.