Hybridizing Learning and Model-Based Planning for Active Perception
Abstract
Project AbstractThe objective of the work in this proposal is to hybridize model-free Machine Learning (ML) withmodel-based planni"ng, control, and state estimation to enable active perception for autonomousvehicles operating in complex environments and in compl""ex mission scenarios. Active perception,meaning real-time decision-making and control for task-driven sensing and state estimation,""provides a powerful force multiplier for situational awareness at all scales from tactical mappingto basin-scale monitoring, incre""asing temporal/spatial resolution and providing robustness.To achieve these capabilities, machine learning methods will provide key"" insights enabling thehigh-level autonomy needed to handle complex stochastic environments, whereas model-basedalgorithms will bri"ng formal robustness needed for reliable mission execution. Our framework decomposesthe decision making hierarchically such that machine learning-based and model-basedmethods are applied where they are most effective via a systematic integration within the hierarchy.Technical Approach Summary. The key technical challenge of the work here is making decisionsfor active perception in highly dynamic operational environments while ensuring reliability and robustnessin achieving mission objectives. We address this challenge" by developing a mathematicaland algorithmic framework to hybridize machine learning with model-based planning, control,and state" estimation to enable active perception in complex data-rich environments. Active perceptionis fundamentally challenging due to the" strong coupling between data acquisition, control,and state estimation. The proposed hybridization inherits the agility and broad" applicability ofmachine learning methods and the robustness of model-based methods: Learned policies processhigh-volume sensed data and handle interactions with the environment to generate directives formodel-based controllers to execute robustly at the vehicl"e level. To facilitate the integration of thetwo approaches, constructed data sets for the learning methods will be built using the"" model-basedmethods, and the resulting capabilities of the learned methods will inform further model-based dataconstruction and wi""ll facilitate generation of the high-level commands that must be executed, inreal-time, by the model-based algorithms. The combinat""ion of the methods will result in provable,robust performance guarantees for fielded single and multi-agent perception systems.Ant"icipated Outcomes. We will apply our methodology to autonomous air and ground vehiclesperforming targeted area search in regions of" limited prior knowledge, and limited or denied GPSavailability, e.g., exploring a dense urban area or the inside of a building. Sp""ecifically, we considerquadrotor air vehicles with limited onboard computational capabilities and a heterogeneous mix ofextrinsic"" sensors. The air capabilities are complemented by ground robots with longer endurance,larger sensor loads, and greater computation"al and communication capacity. Demonstrated resultsfor this system will be immediately translatable to current and planned Naval systems.Future Naval Relevance. This research maps directly onto current and future fleet operationsincluding tactical battlefield r"econnaissance, fleet protection and antisubmarine warfare, harborpatrol, and single or multi-UAV (Unmanned Aerial Vehicle) swarms f""or air superiority. Activeperception is of particular interest for autonomous subsea (e.g., the Forward Deployed Energy andCommuni""cations Output (FDECO) initiative) and small scale air vehicles (e.g., the Low-Cost Unmannedaerial vehicle Swarming Technology (LOC"UST) system) where inherent power and communicationsconstraints limit the capabilities of individual vehicles. The hybridization of learningand model-based methods for advanced perception will provide translational capability improvementfor these applications and many others within ONR and DoD.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jun 09, 2017
- Source ID
- N000141712623
Entities
People
- Kristi Morgansen Hill
Organizations
- Office of Naval Research
- United States Navy
- University of Washington