Quantifying the Effect of Friction Coefficients on the Sim-to-Real Gap in Mobile Robot Path Tracking Simulations on Varied Terrain

Abstract

In the quickly developing field of robotics, there is an urgent and growing need for high-quality and reliable simulations. While robotics simulators are used commonly by roboticists, there is a lack of knowledge about how accurate these simulations are, or how simulation parameters such as the friction coefficient may impact the sim-to-real gap in practical use cases.

To support a more nuanced and resource-effective application of simulations, this thesis presents the validation of several mobile robot path-tracking simulations in the Gazebo robotics simulator using the ODE physics engine and the default pyramid model of friction. Three validation experiments are designed and carried out, testing the robot on varied terrains in real and virtual worlds, comparing simulated behavior to behavior of the real system for each experiment. In the real world testing, mixed reality is used to isolate effects of inadequate robot physics modeling from the effects of inadequately modeled sensors in the path tracking process. The relationship between the friction coefficient set in simulation and the sim-to-real gap of the system is explored by quantifying the sim-to-real gap with key validation metrics of average cross track error, navigation time, and failure rate. Data show a strong insensitivity to the friction coefficient for tests on flat ground and a moderate to high sensitivity to the friction coefficient on inclined and bumpy terrains, especially for lower friction coefficient values. Data also show a significant inconsistency in simulation results for identical experimental trials, suggesting that the simulation model is not robust in tolerating minor variations in the input and/or minor variations in the solution process itself.