NSF Award
2341183
Technological advancements have led to a proliferation of robots using machine learning to assist humans in a wide range of tasks. However, despite the strengths of approaches based on deep learning they have several shortcomings that leave them vulnerable to exploitation from adversaries. In addition, the computational, financial, and environmental cost incurred to train these discriminative models can be quite immense. Alternatively, hybrid methods that combine (less complex) deep learning with other probabilistic techniques can provide more robust and adaptive learning. Unfortunately, these probabilistic techniques tend to suffer from long run times and high computational complexity. This project aims to develop new approaches for hardware acceleration of these probabilistic techniques across a range of robotics applications. These approaches are intended to pave the way for the design of autonomous robots that can sense, perceive, and act in real time in a range of natural human environments, and in a very energy-efficient manner. The proposed work has the potential to enhance human quality-of-life by enabling robots to dramatically expand the range of tasks they can complete. The project also includes curriculum development for an interdisciplinary course in robotic design aimed in part at getting a broader range of students interested in computing and hardware design of robotic systems.<br/><br/>A long-term goal is to reach the point where mobile robots can compute all information needed for perception on-board and in real time. This project focuses on exploiting the complementary properties of deep learning and probabilistic inference for making perceptual decisions, where the weaknesses of one can be addressed by the strengths of the other. The researchers are investigating various algorithmic and hardware-acceleration approaches that provide effective robot perception in unstructured, natural environments in real time and at efficient energy cost. In particular, the research is aimed at goal-directed robot manipulation within a confined embedded system under limited hardware and power budgets. The focus is on probabilistic algorithms such as Bayesian inference that may be incorporated with neural network methods. The project proposes three main tasks: a) accelerating graph-based Bayesian inference in hardware, b) constructing a general-purpose library of optimized hardware modules for accelerating robot-oriented algorithms, and c) using the hardware library to develop new algorithms for robot perception.<br/><br/>This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.
