Abstract
The interest in autonomous mobile robots (AMR) is fast growing in the private, military, and commercial sectors for its promise to revolutionize key components of many industries, such as logistics, structural inspection and transportation. One topic that is not as well studied by academia is the problem of motion planning for AMR in occluded environments. Many practical sensor modalities for AMR are limited to line-of-sight measurements, meaning that detected obstacles (such as a static wall) may occlude the presence of other obstacles from the robot (such as a person moving in a hallway). Occlusions introduce uncertainty into the motion planning problem, as the occluded region may or may not hide obstacles to avoid. This uncertainty is multimodal and highly unstructured, making it a unique challenge towards run time navigation.
This dissertation focuses on the problem of creating an occlusion-aware motion planning policy for AMR. In particular, this work casts the occlusion-aware navigation problem as an optimization problem in which a carefully selected cost function incorporates the desired occlusion-aware behavior, in addition to the usual go-to-goal and obstacle avoidance behavior. Throughout this work, various approaches to casting this optimization problem are explored. First, a visibility-based approach is developed that seeks to approximate the area behind occlusions and actively minimize this area via a Model Predictive Controller (MPC). This results in occlusion-aware motion that requires no pre-training and little overhead to implement within a typical autonomy stack. Second, the occlusion-aware navigation problem is cast as a risk-aware motion planning problem in which occlusion-related uncertainty introduces some element of risk towards the motion of the ego robot. Emphasis is placed upon data-efficient run time learning of common risk metrics, including Value at Risk (VaR) and Conditional Value at Risk (CVaR), requiring relatively fewer datum than other machine-learning approaches. Minimizing this risk amounts to increasing visibility around occlusions, showing how occlusion-aware navigation may be an emergent behavior from risk-aware policies. Additionally, techniques for run time optimizing over these machine-learned risk functions are developed, including a trajectory generation approach based on Model Predictive Path Integral (MPPI) control theory. Lastly, this work investigates techniques for constructing these risk-sensitive policies from data collected at run time. We accomplish this using a Quantile Temporal Difference (QTD) learning algorithm to develop a risk-sensitive policy from direct experiences of the ego robot, as well as an additional approach that infers a risk-sensitive metric from observations of dynamic obstacles near occlusions. The theoretical and practical implications of these approaches are discussed, and these techniques are validated through extensive simulations and proof-of-concept experiments, both inside and outside a controlled lab setting.
