Continuous Mission Execution over Infinite Time Horizon
Name
lee-ianl-meng-eecs-2026-thesis.pdf
Size
10.12 MB
Format
Adobe PDF
Checksum (MD5)
fa06a2d10aff547f82156425a1b6021e
Author(s)
Lee, Ian
Advisor(s)
Williams, Brian C.
Date Issued
February 2026
Publisher
Massachusetts Institute of Technology
Abstract
Autonomous missions are increasingly expected to operate for extended durations, lasting potentially decades in deployed environments. When missions may run indefinitely, there are two challenges with operation. First, the full extent of the plan is too long to reason over at once. Secondly, the duration of actions performed over such a long duration cannot be accurately predicted, and must instead be modeled as a distribution. To address the first problem, we incrementally plan and execute online over a receding horizon. To address planning online with probabilistic duration, we must have an efficient risk-bounded planning method, so that planning can stay ahead of execution over the receding horizon while keeping temporal risk—the chance of missing deadlines—in bound. This thesis presents algorithmic contributions that address both. First, we contribute CRTE*, an online algorithm that makes execution decisions continuously over a receding horizon. Existing approaches assume the full plan is given before execution and react only to event outcomes, not to newly arriving plans. This leaves open how to transition between plans while maintaining consistency with already-committed execution times. CRTE* addresses this by preserving execution history across plan transitions and replaying it selectively through each new plan’s constraints. We build an execution system based on CRTE* and demonstrate it on a simulated planetary rover mission. Secondly, we contribute a conflict extraction algorithm for Rubato, a probabilistic temporal controllability checker, so it can be used as a subsolver for online conflict-directed planning. Rubato converts probabilistic scheduling constraints to bounded ones using a budget of allowed risk, checking feasibility through a generate-and-test procedure over risk allocations. When this procedure fails, infeasibility is detected only indirectly—through failures of induced bounded networks under particular allocations—making it unclear which original probabilistic constraints are responsible. Our algorithm traces these failures back to the original constraint level, enabling upstream conflict-directed search for more efficient risk-bounded planning.
MIT Department
Massachusetts Institute of Technology. Department of Electrical Engineering and Computer Science
Terms of Use
In Copyright - Educational Use Permitted
Copyright retained by author(s)
Persistent DSpace Link
