Hybridizing Learning and Model-Based Planning for Active Perception

From u-t-autonomous.info
Jump to: navigation, search

Technical point of contact: Marc Steinberg, Office of Naval Research
Period of activity: 2017-2020
Collaborators: Kristi Morgansen (PI, Univ of Washington), Behcet Acikmese (Univ of Washington), Sergey Levine (Univ of California, Berkeley)

Overview of the Project

The objective of this project is to hybridize model-free machine learning with model-based planning, control, and state estimation to enable active perception for autonomous vehicles operating in complex environments and in complex 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 mapping to basin-scale monitoring, increasing temporal/spatial resolution and providing robustness. To achieve these capabilities, machine learning methods will provide key insights enabling the high-level autonomy needed to handle complex stochastic environments, whereas model-based algorithms will bring formal robustness needed for reliable mission execution. Our framework decomposes the decision making hierarchically such that machine learning-based and model-based methods are applied where they are most effective via a systematic integration within the hierarchy. caption

Active perception is fundamentally challenging due to the coupling between data acquisition, control, and state estimation. Hybridization inherits the agility and broad applicability of machine learning methods and the robustness of model-based methods: Learned policies process high-volume sensed data and handle interactions with the environment to generate directives for model-based controllers to execute robustly at the vehicle level.