Existing robotic arms have limited or no ability to avoid collisions with their environment due mainly to the lack of a suitable sensing system. A collision avoidance capability should be incorporated into every robot so that injuries to people and damage to equipment from collisions are prevented. Important applications that could benefit from robot collision avoidance include:​ manufacturing, robot-assisted surgery, robotic handling of hazardous waste, and personal robots.

Creating a full-coverage, fast, reliable and cost effective sensing system for sensor-based robotic arm collision avoidance is a challenging problem. Capacitive sensors were selected based on their promising potential. Capacitive sensors have the limitations of nonlinearity and being influenced by the environment. In this thesis, their sensing behaviour, and solutions to these limitations, were investigated.

A forward model predicts the capacitance for a given electrode geometry. The conventional method, Method of Moments (MoM) and Finite Element Method (FEM) were investigated and compared. The MoM demonstrated that the fringing electric field ignored by the conventional forward model is significant for the robotic arm application due to the relatively large ratio of electrode gap to electrode area. Two forward modeling cases were simulated by writing macro code for a commercial FEM package. The first consisted of two parallel cylindrical robotic arms. The second consisted of two cylindrical shell electrodes wrapped around a pair of robot links that rotated relative to each other. The results for this case were compared with experimental results. The FEM results were a poor predictor of the experimental results. The failure of the FEM model to include the true environmental conditions (e.g. air humidity and surrounding electric fields) is the most likely cause of its inaccuracy.

An inverse capacitance model outputs the electrode geometry for a given capacitance. In this research the desired geometric output was the seven robot link pose variables, (x, y, z, q_x, q_y, q_z, q₀), describing the position and the orientation of the link of a robotic arm. A Cerebellar Model Articulation Controller (CMAC) neural network was chosen for the inverse modeling based its ability to model nonlinear behaviour and its efficiency. One CMAC network was trained for each pose variable. The sensor was built using capacitance sensing circuit and a multiplexor board with the potential for 16 by 16 electrode combinations. Note that an n by n combination produces n² separate capacitance values.

For the inverse modeling experiments, four aluminum foil electrodes were mounted on a CRS-F3 robotic arm and four aluminum foil electrodes were placed on a wooden box used to simulate a second stationary robotic arm. A pair of reference electrodes was mounted on the back of the CRS-F3 arm. This reference measurement was used to normalize the measured capacitances in order to minimize environmental effects. The normalized capacitance data were used to train and test the CMAC neural networks. The CMAC learning factors were dynamically changed to reduce the training errors. A new fuzzy logic approach was developed that allowed the range of the CMAC input data to be increased without significantly increasing the training error. After evaluating eleven combinations of electrodes, it was determined that only the 3 by 3 and 4 by 4 combinations converged with small training errors. Three methods were used to analyze the CMAC testing errors:​ comparison plots, error plots and error metrics. Over a 15 cm range, pose variable y had maximum absolute errors of 2.1 mm for the 4 by 4 electrode combination and 7.2 mm for the 3 by 3 electrode combination. For the 4 by 4 combination the maximum relative errors were less than 3% for the x, y, and z variables, and less than 15% for the quaternion variables. For the 3 by 3 combination, these values increased to 13% and 20%, respectively. The larger relative errors for the quaternion variables were due to their smaller ranges of variation.

Using the same hardware, a simple collision avoidance system was implemented using one pair of electrodes to detect the potential collision between a robotic arm moving in the vertical plane and a second stationary robot. The robot was shown to successfully avoid the potential collision and then continue its motion.