Abstract:
Given the increase in scale and complexity of robotics, robot application development is challenging in the real world. It may be expensive, unsafe, or impractical to collect data, or test systems, in reality. Simulation provides an answer to these challenges. In simulation, data collection is relatively inexpensive, scenes can be procedurally generated, and state information is trivially available. Despite these benefits, the use of simulators is often limited to the early stages of application development. In this work, we take steps to close the gap between simulation and reality, for Lidar simulation.

We adopt the perspective that the eventual purpose of a simulator is a tool for robot application development. Our framework for sensor simulation consists of three components. The first is sensor modeling, which describes how a sensor interacts with a scene. The second is scene generation, needed to construct simulated worlds corresponding to reality. The third is simulator evaluation, based on comparing real and simulated data. We formalize the intuition that application performance must be similar in simulation and reality, using an application-level loss. Our framework is broadly applicable to simulating sensors other than Lidars.

We instantiate our framework for two domains. The first domain is planar Lidar simulation in indoor scenes. We construct a high-fidelity simulator using a parametric sensor model. We show how application development paths for our simulator are closer to reality, compared to a baseline. We also pose sensor modeling as a case of distribution regression, which leads to a novel application of a nonparametric method that adapts to trends in sensor data. The second domain is Lidar simulation in off-road scenes. Our approach is to build a library of terrain primitives, derived from real Lidar observations. These are shown to generalize, resulting in an expressive simulator for complex off-road scenes. For this domain as well, we quantitatively demonstrate that our simulator is better for application development, compared to a baseline.

Our work suggests a generic approach to building useful simulators. We view them as predictive models, and perform thorough tests on real data. We evaluate them with an application-level loss, which supports their greater use in the development cycle.

More Information

Thesis Committee Members:
Alonzo Kelly, Chair
Martial Hebert
Michael Kaess
Peter Corke, Queensland University of Technology