Neuromorphic control of dynamic systems
Author(s)Singh, Prince, Sc. D. Massachusetts Institute of Technology
Massachusetts Institute of Technology. Department of Aeronautics and Astronautics.
MetadataShow full item record
Arguably, the agility of a robotic system is dictated by the speed of its processing pipeline, i.e., the speed of data acquisition plus data processing from a robot's on-board vision sensors. Specifically, one ideally hopes that this pipeline offer fresh measurements at a high temporal resolution with low-latency in a computationally-cheap manner for efficient control. This desirable situation may be hard to come by for majority of the current vision-based robotic applications that rely on the traditional CCD-/CMOS-pipeline, as one would be in search for traditional cameras that offer a high sampling rate (thus, high temporal resolution) whose potentially redundant (thus, not fresh) and synchronous output must be processed with low-latency in a computationally-cheap manner. For instance, processing the synchronous series of conventional camera images, which embed possibly redundant levels of intensities may greatly hinder the fast reaction times required by robots while expending power. This issue warrants the need for faster sensors in order to truly address the underlying perception problem for high-performance systems that need to operate under power constraints. To this end, we capitalize upon the merits of a recently introduced biologically inspired and computationally-cheap alternative to traditional cameras-called Neuromorphic Vision Sensors whose pixels independently and asynchronously (thus, high temporal resolution) fire, in the order of micro-seconds (thus, low-latency and high temporal resolution), a stream of non-redundant (thus, fresh) brightness changes represented as binary numbers (±1), termed retinal events, based on a trigger condition that is defined on a logarithmic scale. These properties offer a faster processing pipeline and hint that the Neuromorphic sensor would be a promising candidate to facilitate high-speed robotic applications. However, existing computer-vision based algorithms designed for processing periodic measurements cannot be directly adapted to process retinal events, as these are fired aperiodically, and are ambiguous since they are binary. As an additional challenge, in practice, many retinal events are misfired due to the presence of underlying sensor circuitry noise (not associated to physical brightness changes in the environment) and we term these as spurious events. The merits and operational constraints of this vision sensor mandates the development of a corresponding control-theoretic setup. Thus, the contributions of this dissertation are twofold: 1) to design a control algorithm that processes de-noised retinal events to facilitate a prescribed control task, and 2) to propose a de-noising procedure that mitigates the effect of spuriosity in retinal events. The first part of this dissertation, investigates the problem of controlling (i.e., stabilization and regulation) a Continuous-Time Linear Time Invariant (CT-LTI) system using retinal events generated from an idealistic model of a Neuromorphic Vision Sensor, which is an instance of a broad family of signal change detection sensors frequently encountered in practice. The contribution is to present a novel control design procedure that stabilizes and regulates a hybrid system, consisting of the CT-LTI system and the discrete-event signal change observation model, to a desired set-point. Moreover, the set of thresholds (sufficient conditions) for the given system to fulfill the prescribed control task is provided. The proposed controller is then extended to handle the case of noise in both the system dynamics as well as the observation model; thus, accounts for spurious events in this setting. The second part of this dissertation proposes a de-noising algorithm-Spuriosity Filter (SF)-and is motivated by the practical need to reduce spurious events whilst working with general observation models. The construction of SF is based on the fundamental lack of spatial correlation between spurious events and the algorithm trades off pixel resolution to produce a cleaner event stream on larger spatial scales by seeking a form of 'consensus' between neighboring pixels. At the core of our analysis lies a formal equivalence relation, defined as a means to track brightness, between our filter and a lower-resolution neuromorphic sensor with reduced noise levels. As a consequence of the principled analysis, we highlight important properties that any filter, which processes asynchronous noisy retinal events must respect and have not been accounted for by existing works. The effectiveness of the proposed control-theoretic setup for the illustrative task of heading regulation is illustrated over a range of systems: from numerical experiments to a laboratory testbed.
Thesis: Sc. D., Massachusetts Institute of Technology, Department of Aeronautics and Astronautics, 2018.Cataloged from PDF version of thesis.Includes bibliographical references (pages 127-145).
DepartmentMassachusetts Institute of Technology. Department of Aeronautics and Astronautics.
Massachusetts Institute of Technology
Aeronautics and Astronautics.