Belief-Space Planning for Real-World Systems: Efficient SLAM-Based Belief Propagation and Continuous-Time Safety
Name
frey-kfrey-phd-AeroAstro-2021-thesis.pdf
Description
Thesis PDF
Size
6.67 MB
Format
Adobe PDF
Checksum (MD5)
3a9cc9126cc5be9e8367faf9b08206e8
Author(s)
Frey, Kristoffer M.
Advisor(s)
How, Jonathan P.
Date Issued
June 2021
Publisher
Massachusetts Institute of Technology
Abstract
Uncertainty-aware planning has long been a recurring goal in robotics. By enabling autonomous systems to explicitly reason about their own uncertainty, desirable behaviors that increase observability and ensure robust constraint satisfaction arise naturally from high-level optimization specifications. For partially-observable and under-sensed systems in particular, belief-space planning (BSP) provides a natural probabilistic formulation. Despite significant research attention over the years, a few key challenges have prevented the application of BSP to the real-world systems that would stand to benefit the most, such as SLAM-reliant Micro-Aerial Vehicles (MAVs).
The most fundamental of these challenges is that of efficiently propagating the state belief, particularly under SLAM-based estimation schemes like Visual-Inertial Odometry (VIO). This thesis describes a structureless and consistent approximation for
belief propagation under SLAM, the efficacy of which is demonstrated in the challenging setting of observability-aware planning for VIO.
A key attraction of BSP is the ability to specify constraints on the total probability of failure – however, actually encoding these constraints within practical optimization schemes remains a challenge, particularly for physical systems, which evolve continuously in time. General-purpose Monte-Carlo methods can be used to accurately assess failure rates, but these are cumbersome to optimize against, while more convenient “direct” estimates are based on discrete-time simplifications and fail to meaningfully constrain the full continuous-time risk. To address this gap, a novel risk estimate is derived directly in continuous-time, providing a principled, lightweight, and convenient means of ensuring probabilistic safety for real-world systems. Together, these contributions enable online, risk-constrained BSP for a large class of systems of widespread practical interest.
MIT Department
Massachusetts Institute of Technology. Department of Aeronautics and Astronautics
Terms of Use
In Copyright - Educational Use Permitted
Copyright retained by author(s)
Persistent DSpace Link
