Sensing, Machine learning, and Robotics Technologies Laboratory
The mission of the Sensing, Machine learning and Robotics Technologies Lab (SMARTLab) is to create new knowledge, pursue engineering applications, and train future-generation professionals in signal processing, machine learning, robotics, and automation systems.
Machine Learning, Signal Processing, and Control Theories and Methods
Computer Perception and Visual Understanding
Signal Processing for Next-Generation Wireless and Data Storage Systems
Distributed Sensing, Estimation, and Optimization
Medical Imaging, Brain-Computer Interface, and Assistive Technology
Autonomous Robotic Systems
Radar, Sonar, and Remote Sensing
Analysis and Design of Complex Networked Systems
Cyber-Physical Systems, Network Systems, Control, Optimization, and Artificial Intelligence
This work develops a detection algorithm for joint multiuser detection in asynchronous wireless networks. Joint multiuser detection of multiple digitally modulated signals improves performance over a conventional receiver that separately performs user-by-user detection. However, one major problem is the asynchrony of the different signals transmitted from different users. This asynchrony can arise when multiple signals have slightly different baud rates. If not efficiently dealt with, this asynchrony degrades the joint detection performance compared to the performance in the corresponding synchronous network. This work aims to develop a joint detector that internally deals with the existing asynchrony of the incoming signals.
This work develops a multitrack extension of M&M TED. M&M TED is originally derived for the simple case of synchronizing a single waveform to unknown timings and therein it fits within a one-dimensional detection problem. However, for joint detection of multiple digitally modulated signals whose signaling rates are asynchronous, a multitrack extension of M&M TED is required. This multitrack TED will include the cross-talk in estimating the timing error of each of the signals.
Most powerful and widely used machine learning tools used in BCI, such as Neural Networks, require extensive training data while suffering from over-fitting problem. This work plans to develop adaptive EEG signal processing algorithm aiming to minimize the long and problematic training sessions of each BCI user while maximizing the robustness of the algorithm in response to different settings in which the algorithm is used.
An Integrated Robot and Wearable Sensor Approach
Funded by NSF (2019-2022)
The project will develop an integrated autonomous system that consists of a mobile robot and smart insole sensors to assist older adults to independently live in their own homes and interact with their communities. The system guides individuals in regular walking exercises, autonomously assesses gait states, and provides real-time personalized feedback to engage older adults into the exercise. The robot will also be used to connect older adults with family and friends through a virtual connection. The project team will evaluate the system at a senior center in New York City using objective and subjective performance criteria measuring older adults’ experiences with the system.
Funded by NSF (2018-2021)
The project develops a robot motion planner with human-like navigation features for improved pedestrian/mobile-robot navigation through crowded and unconstrained environments. Current mobile robot motion planners are inadequate because current models of pedestrian dynamics do not fully capture the complexity of human motion behavior in crowds. The project will use novel machine learning techniques to extract features related to pedestrian behavior from existing datasets and then train a deep neural network to model pedestrian dynamics. The project will also study socially normative behaviors and criteria associated with robot navigation. The project contributes to human-robot interaction, and may result in public safety applications such as emergency evacuations and crowd planning/management.
Funded by NSF (2015-2019)
Having well-designed robotic systems to assist people in public crowd environments such as shopping malls, museums, and campus buildings benefits society economically. More important, in life-threatening emergency situations, robot-assisted evacuation could save lives by reducing congestion and preventing crowd stampede, which is the most disastrous forms of collective human behavior induced by escape panic in emergency. The project exploits human-robot interaction for robot-assisted evacuation. Adaptive dynamic programming and deep reinforcement learning algorithms are developed for robots to interact with pedestrian flows for effective pedestrian regulation.
Funded by NSF (2016-2020)
This project considers passive radio frequency (RF) sensing that employs wireless communication signals as illuminators of opportunities (IOs) to detect, locate, and track objects of interest. Such capabilities are useful for a wide range of applications, e.g., indoor localization, health monitoring, vehicle tracking, and many more. Passive RF sensing has many advantages over its active counterpart: no dedicated transmitter and RF pollution free; covert operation; order of magnitude cheaper to build, deploy and operate; and the ability to simultaneously access several IOs to obtain multiple views and spatial diversity of the surveillance area. Despite the advantages, there are fundamental technical issues that need to be investigated for passive RF sensing. This project aims to develop novel signal processing techniques for passive RF sensing by taking into account impairments such as noisy reference, direct-path interference, and multi-path clutter, which are inherent in passive systems.
Adaptive Radar Signal Detection with Integrated Learning and Knowledge Exploitation
Funded by AFOSR (2016-2020)
This project addresses the fundamental problem of weak signal detection in strong clutter for airborne radar. Due to high-speed platform motion, the clutter often exhibits excessive angle-Doppler spread and non-homogeneity in range and azimuth, which renders ineffective conventional covariance matrix based detectors that require ample homogeneous training data for clutter mitigation. The broad research objective is to develop a robust Bayesian learning and detection framework which can exploit prior knowledge of the clutter to compensate for lack of training data and, meanwhile, is invulnerable to erroneous or outdated knowledge via joint learning and validation using the test data.
Vast amount of images are routinely collected by ubiquitous intelligence, surveillance, and reconnaissance (ISR) systems. In many mission scenarios, what may appear in these images is not exactly known in advance. Therefore it generally requires intensive manual analysis by well-trained specialists to interpret the content and identify potential target or threat. In this project we attempt to develop an automated system for unsupervised object discoveries and descriptions. The task of object discovery involves visual object detection, recognition and localization within any given image frame. Subsequently, the task of object description provides high level semantic description that ultimately can capture the visual, relational, behavioral and functional properties of the discovered objects.
Bridging visual and textual semantics has always been a major challenge in multi-modal machine learning. One particular research problem we have been addressing is sentiment analysis on mixtures of short textual and visual messages commonly found in social media. In this project we are first developing two separate novel methods for analyzing images and text messages, which share a common structure consisting of multi-scale representations, metric learning and manifold embedding. To establish a linkage between these two modalities, we will further introduce a new visual sentiment descriptor termed “bag of affective words” (BoAW), and an efficient way of estimating the conditional probabilities of adjectives and nouns pairs (ANPs) given an image represented by it BoAWs. These ANPs are then combined with multi-scale textual features in the form of “set of sentiment parts” (SSP) in the joint sentiment analysis.
We have a set of promising preliminary results, that provide a closed-form distributed solution of the optimal (in the sense of lp induced norms) ordering decisions that control inventory positions at each echelon in the supply chain. The remarkable feature of the aforementioned decision policies is that (under mild technical assumptions) they are able to virtually eliminate the so-called “bullwhip” effect from the behavior of the chain. Existing ordering policies are unable to cope with the propagation (and in most cases the amplification) of the “orders for products” upstream the supply chain, resulting in increasingly larger inventory variations - a highly undesired phenomenon dubbed the Forrester (or bullwhip) effect. In our method, the effect of the “orders of products” (assimilated to disturbances in the control parlance) attains optimal attenuation. Our main technical ingredient is a tailored adaptation of very recent methods borrowed from distributed control.
Large-scale optimization is the central scientific factor in a vast array of technical applications, including machine learning, managing the power grid or today’s supply chains. Because of the enormous size of the data, the existing solutions mostly rely on first order or proximal point methods. It has been shown recently that such iterative algorithms may be assimilated to discrete-time controllers while the objective function to be optimized may be viewed as a plant in the presence of model uncertainty. We are leveraging on our preliminary results from distributed optimal control, in order to come up with high performance numerical optimization algorithms for large scale problems.
Funded by NSF (5-year CAREER Award)
A certain class of oscillatory behaviors, popularly known as the butterfly effect or the bullwhip effect are quite habitual to large scale networks of dynamical systems, such as electrical networks, supply chains or transportation networks. They essentially describe as tiny perturbations (or fluctuations) in the functional parameters occurring in one area of the network, start to amplify while propagating through the dynamics of the network and ultimately result in extremely large deviations from the operational parameters in other parts of the network. This undesired phenomenon is very common and difficult to avert in important practical applications. In the case of a string of road vehicles (vehicles platooning) it translates in the common stop-and-go (slinky or accordion effect) noticeable on a busy highway and which is extremely detrimental to the highway traffic. Our preliminary results contain a design method for an optimal cruise control system, encompassing a comprehensive solution to the automated platooning problem that is fail safe and viable from all engineering standpoints.
A group of subsystems/agents that collaborate autonomously through physical interactions and/or communications (forming a networked system/multi-agent system) could achieve objectives that are beyond the capability of an individual subsystem/agent. Moreover, such networked systems could offer better efficiency, fault tolerance, and flexibility due to their distributed operations, parallelized executions, function redundancy, and reconfigurability in the face of changes. Most importantly, many critical real-world systems, like power networks, supply chain networks, platoons, etc., can be modeled as networked systems. Consequently, there has been a steady growth in research interest in networked systems both from academia and industry.
One of the significant concerns in engineering such networked systems is the design of the cooperative controllers and the interconnection topology while guaranteeing compositionality and scalability. Here, compositionality implies that when certain subsystems are added to or subtracted from the existing network, there is no need to redesign the controllers and the topology globally, but only need to modify the affected components locally. On the other hand, scalability describes how the properties of the networked system, e.g., passivity indices, stability margins, etc., scale with respect to the system dimensions. When these two properties are retained, the subsystems can work cooperatively as a team and are free to join or leave the existing networks without any influence on the properties of the entire network. Besides, for large-scale networks, not only the controllers but also the topology plays a crucial role, as improper topologies may lead to performance deterioration, high delays, high communication costs, and even instability. In this regard, topology synthesis techniques are usually required to obtain an optimal topology with the least communication cost and the optimal system metrics, e.g., a robustness measure with respect to uncertainties, security-aware metrics, and stability margins.
Based on the above observations, in this project, we study the problem of compositional and scalable controller and topology co-design. In our prior work, we have developed several critical fundamental results to address this problem in a more generalized setting. The ongoing research explores establishing more specialized results necessary to address this problem for real-world networked systems like vehicular platoons, power networks, and supply chain networks.
Cooperative multi-robot systems (CMRSs) are used in many applications such as assembly lines, delivery systems, transportation systems, maintenance systems, and persistent monitoring systems. This research project focuses on persistent monitoring (PM) applications of CMRSs, which include applications such as patrolling, surveillance, data collection, particle tracking, and distributed sensing/estimation. The goal of PM is to use a CMRS (i.e., a team of agents) to optimally monitor an environment containing several dynamically updating (state) points of interest (i.e., a group of targets). The main challenge faced by existing PM solutions is their inability to simultaneously handle the underlying multiple objectives, dynamic constraints, and hybrid dynamics required of realistic PM systems without compromising their scalability, optimality, and robustness. Towards addressing this challenge, in our research, we have designed a generic graph-based PM (PMN) problem framework and a corresponding event-driven receding horizon control (RHC) based PM solution framework with many attractive qualities such as being efficient, on-line, gradient-free, and parameter-free. The ongoing research attempts to address several remaining aspects of the said main challenge (e.g., robustness) as well as to extend the proposed PMN problem and RHC solution frameworks via integrating the latest research discoveries on control barrier functions, on-line learning, distributed consensus, and networked control.
Several critical problems arising in (and not limited to) multi-agent systems, machine learning, data mining, and scheduling may be formulated as set function maximization problems subject to cardinality constraints. In such problems, the interested set (objective) functions often have monotonicity and submodularity properties (in other words, they increase with the number of agents but have diminishing returns). Despite its versatility, owing to its challenging nature, almost all existing solutions to this class of submodular maximization problems are based on greedy algorithms. A seminal work on this topic has exploited the submodularity property to prove a (1-1/e) performance bound for such greedy solutions. More recent literature on this topic has been focused on exploiting different curvature properties to establish improved (tighter) performance bounds. However, such improvements come at the cost of enforcing additional assumptions and increasing the computational complexity of evaluating such performance bounds. In our recent work, we have proposed a new performance bound that does not require any additional assumptions and is both computationally inexpensive to evaluate and highly practical. Motivated by these findings, in this project, we explore more general and specialized versions of the submodular maximization problems to derive improved and computationally inexpensive performance bounds. To verify the effectiveness of the established results, we plan to use the widely studied classes of applications: multi-agent coverage control and multi-satellite assignment.