Risk-Aware Planning and Control for Autonomous Systems: Models and Real-Time Algorithms
Abstract
Abstract: The goal of this proposal is to devise efficient algorithms for risk-aware planning and control of autonomous systems in uncertain environments, where threats, dynamic events, and partial knowledge of world models are dominant features. To enable an autonomous system (e.g., an unmanned aerial vehicle) to plan its actions under uncertainty, one needs to design a strategy for a decision maker, rather than an open-loop motion plan. Most of the current approaches aim at optimizing average performance, and appear inadequate to address problems in which events with low probability can have a detrimental impact (e.g., detection by radars, collisions, and loss of communication). In contrast, in this effort we will devise control and planning algorithms that explicitly take into account risk, i.e., increased awareness of events of small probability and high consequences. A fundamental challenge for risk-aware planning and control is the assessment of risk in a dynamic setting, where decisions are made over time. The key novelty of this effort is to leverage recent strides in the theory of time-consistent, dynamic risk metrics developed by the operations research community, to lay the foundation for a rational, tractable, and operationally sound theory of risk-aware planning and control. Specifically, we will model the problem of risk-aware planning and control as a Markov decision process, constrained by time-consistent, dynamic risk metrics. By assessing risk at multiple points in time, such metrics guarantee the time consistency of risk assessments. The advantages of our formulation include its axiomatic justification (in terms of “time consistency”), and its amenability to dynamic optimization due to the compositional structure of dynamic risk metrics. Our research objectives are: • Objective 1. Theoretical foundations: Address the theoretical underpinnings for the application of the theory of time-consistent, dynamic risk metrics to the problem of risk-aware planning and control. • Objective 2. Solution algorithms: Devise exact and approximate solution algorithms (e.g., dynamic programming and real-time sampling-based methods) for risk-constrained Markov decision processes, where risk is assessed according to a time-consistent, dynamic risk metric. The focus will be on algorithms amenable to implementation on embedded systems. • Objective 3. Calibration and validation: Identify, in collaboration with the U.S. Navy, the most appropriate risk metrics for U.S. Navy operations and validate the algorithms on a quadrotor test bed. By placing risk-aware planning and control on a firm theoretical foundation that is compatible with the modern theory of risk, this effort will bring the field of planning and control one step closer to the U.S. Navy s vision of autonomous, adaptive systems that can safely operate in uncertain and unstructured environments. As such, this effort is closely aligned with the ONR Code 31 research program, in particular, the “Machine Learning, Reasoning and Intelligence” program.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Aug 12, 2016
- Source ID
- N000141512673
Entities
People
- Marco Pavone
Organizations
- Office of Naval Research
- Stanford University
- United States Navy