Uncertainty Propagation in Robot Perception

Abstract

Autonomous construction of shelters on Mars, robotic assistants for life-saving surgeries, and home robots integrated into our lives, once viewed as science fiction, will be attainable in the near-future thanks to advancements in robotics. For a robot to reason about its surroundings during these complex tasks, it relies on noisy sensor data. Thus, the algorithms we employ for perceptual understanding should quantify and propagate the uncertainty induced by sensor noise rather than ignore it. Additionally, robots must quantify and propagate uncertainty in real-time under computational constraints that arise due to the inherent hardware limitations of mobile robotic platforms. By quantifying and propagating perceptual uncertainty, robots can learn through their interactions with their environment and use this uncertainty for safe, robust, and reliable decision-making.

Creating these safe, robust, and reliable robotic systems necessary to accomplish these challenging tasks first necessitates the quantification of both proprioceptive (e.g. system state) and exteroceptive (e.g., metric-semantic mapping) uncertainty. Next, by propagating this uncertainty given new sensor data, a robot is capable of learning about its environment.By quantifying and propagating both proprioceptive and exteroceptive uncertainty, robots can plan more effectively and, in the case of failure or collision, can describe the factors leading to failure in a human-interpretable fashion. Quantifying and propagating uncertainty in robotics has been challenging for two reasons. First, both proprioceptive and exteroceptive models in robotics are often high-dimensional. Consequently, algorithms designed to handle such high-dimensional models must be efficient while managing finite computational memory to operate in real-time. Second, real-world systems are nonlinear and their associated models of uncertainty are non-Gaussian, making tractable closed-form propagation challenging.

This dissertation tackles the challenges associated with uncertainty in robotic perception by proposing a framework for uncertainty propagation for both proprioceptive and exteroceptive models. In particular this dissertation focuses on (1) uncertainty quantification and propagation for robotic metric-semantic maps representing exteroceptive sensing and (2) uncertainty propagation over Riemannian manifolds, which play a key role in robotic proprioception. These contributions allow roboticists to take an important step towards not only quantifying but also propagating uncertainty in perceptual data for robotic systems in real-time while operating with limited computational resources.