Active Perception and Learning for Robust Sensing and Control in UAVs
Abstract
The primary aim of this project is to develop new biologically inspired models of sensorimotor control that can be implemented on aerial robots to enable autonomous exploration of unfamiliar environments. We will focus on solutions that rely primarily on vision, so that such robots can be deployed in environments where GPS and accurate maps are unavailable. Our approach makes use of a neural architecture for building models that leverages key principles of neural processing, active perception, and predictive coding.The project comprises four tasks that build upon one another, leading to a progressively enhanced degree of autonomy in the system~s capabilities for perception, on-line visual guidance, and navigation.Task 1 focuses on perceptual capabilities that support the ability to safely and effectively move about in cluttered environments ~ specifically, the estimation of self-motion and object motion. We will adapt our recently developed neural model of optic flow processing for use on board an unmanned aerial vehicle (UAV), and conduct systematic tests of its effectiveness. This effort will also involve an expansion of the model to allow for the detection of non-visual (e.g., inertial, motor) signals and the multisensory processing of visual and non-visual information. The aim of Task 2 is to expand the neural model from Task 1 to include mechanisms that support online visual guidance. This effort will draw upon the principles of predictive coding, which offers a powerful framework for capturing the robustness and flexibility of visual control in humans and other animals. Our predictive-coding-compatible models should be more tolerant to interruptions of sensory input (e.g., due to occlusion or sensor malfunction) and better able to intelligently adapt their behavior to variations in task demands (e.g., the need arrive at a location as quickly as possible). In Task 3, we will develop a novel method for using machine learning to extract strategies for on-line visual guidance from experienced human drone pilots. This effort will be distinct from previous attempts to use learning-from-demonstration insofar as the extracted control strategies will be concordant with the tenets of predictive coding, and therefore compatible with the neural model of on-line visual guidance from Task 2. Taken together, Tasks 2 and 3 give us a procedure for learning human-like strategies for visual guidance that can be implemented on board a UAV. The proposed research will be conducted by a multi-disciplinary team with expertise in all the relevant areas, including vision science, neuroscience, biological neural modeling, bio-inspired aerial robotics, and control systems engineering. Successful completion of the project will give aerial robots intelligent sensorimotor capabilities, greater autonomy, and enhanced situational awareness to complete missions with less human supervision.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jul 10, 2018
- Source ID
- N000141812283
Entities
People
- Brett R Fajen
Organizations
- Office of Naval Research
- Rensselaer Polytechnic Institute
- United States Navy