Whisker-Inspired Sensors for Unstructured Environments

Abstract

Robots lack the perception abilities of animals, which is one reason they can not achieve complex control in outdoor unstructured environments with the same ease as animals. This thesis focuses on increasing the perception ability of the bio-inspired whisker sensor, an underdeveloped sensor for robotic perception in contact mapping on terrestrial robots and airflow during a drone flight. Prior work into engineered whisker-inspired sensors demonstrates that the whisker's high compliance, sensing in low visibility, and 3D perception abilities are assets for robotic sensing. Here, we enhance the benefits of prior whisker-inspired sensors by developing sensors and algorithms that are robust to three confounding signals we consider as likely to occur in unstructured environments: 1. objects with compliance 2. simultaneous signals (e.g., wind when trying to sense contact) 3. wind during flight. Through our development of the sensors and algorithms, we found key insights that could be applied to future engineered whisker sensors.

The most popular contact mapping algorithm assumes contact does not occur on compliant surfaces, surfaces that are prevalent in real-world environments. We demonstrate that a sensor that compares the output of this contact point estimation algorithm with a second contact point estimation algorithm can quantify the divergence between the two estimates. This divergence metric indicates the likelihood that the no-compliance assumption has failed. The metric provides insights into the whisker sensor's environment and allows a robot platform to rely on precision estimates until there are indications of algorithm failure (Chapter 2).

In a sensor called WhiskSight, we demonstrated how global stimuli (airflow, body motion) affect an array of whiskers uniformly while local stimuli (contact) affect each whisker in an array differently. This distinction allowed an algorithm to distinguish the source of whisker response between contact, airflow, and body motion occurring simultaneously, eliminating prior requirements for the whisker sensor to know the stimuli it measured (Chapter 3).

The WhiskSight sensor provided the key insight into our work on flow sensors for quadrotors: comparing signals from an array of whisker sensors could be more informative than averaging single sensors. In the two flow arrays we created we used different methods to create whiskers with asymmetric sensitivity to flow based on flow heading. When we compared the signals of all the whiskers in the array, the signals illuminated causal flow measured on a quadrotor. In Chapter 4, the whisker sensors' cross section was asymmetric, and varying the orientation of the whisker led to different response strengths, improving flow heading prediction accuracy by 19$%$.

In Chapter 5 and 6, the asymmetry was generated using a densely packed whisker array. This array made it possible to distinguish two simultaneous flow headings, which is important as quadrotors experience both motion-inducedd drag and wind. With a 2 x 2 array of densley packed whiskers we could predict the heading of a second flow source with a Root Mean Square Error of SI{5.3}{degree}.