The scale, complex structures and often harsh environments of assets such as bridges, powerlines, offshore wind turbines, and oil and gas platforms and pipelines, pose fundamental challenges to planning and control. The standard approach is to break the problem down into:
- A high-level mission-planning component, based on very simplified models of robot dynamics, followed by
- A motion-planning component, using more accurate models of nonlinear dynamics but ignoring issues such as sensing and wind/current disturbances, followed by
- A control component, which tries to achieve the motion plan despite these uncertainties.
However, such a separation is not feasible when the robot’s operation space is highly constrained, which is often the case when the robot operates in close proximity to the assets, or when uncertainties are too significant to ignore. In such scenarios, most plans generated based on simplified dynamics will likely be infeasible for the control component to follow. Therefore, an integrated approach is necessary. To develop an integrated approach, we need to answer fundamental questions in three areas:
Scale and computational complexity of planning under uncertainty.
Mathematically speaking, planning for information-gathering in partially known environments can be formulated as a problem of planning in belief space. Even for relatively simple scenarios, however, this problem is infinite-dimensional, highly nonlinear, and computationally intractable. To resolve this problem, we will: formulate tractable finite-dimensional representations of the problem; investigate the kinds of approximations that can be made without sacrificing quality; and develop ways to incorporate the properties of complex sensing systems into the high-level planning process.
Efficient and accurate predictions of uncertain dynamical systems.
The proximity to assets of robots that are disturbed by external forces (such as wind and currents) necessitates motion planning with safety guarantees. However, most current approaches become meaningless when uncertainty is large, due to the difficulty in accurately bounding possible system behaviours. To address this issue, we will: develop techniques to ensure safety in a meaningful way when the magnitude and uncertainty of external disturbances are substantial and fast replanning is required.
Adaptation to change and integration of physics with learning.
A long-term goal of asset management is for robots to inspect and monitor assets autonomously. This requires robots to be adaptive – able to learn from past experience and incorporate this knowledge to improve subsequent planning and learning. We will develop methods to efficiently combine planning, learning and control in a scalable manner without sacrificing safety or the mission objectives.