Heterogeneity, Diversity and Resilience
in Multi-Robot Systems

August 15-16, 2016
Westin Arlington, VA

Abstracts

Group 1: Resilient Autonomy

Gaurav Sukhatme - Resilient autonomous robot teams in dynamic environments

We are engaged in a long term program of study of robot teams that can function autonomously in highly dynamic environments. We are particularly interested in resilient teams in such settings. Our focus in on resilience to a. significant and poorly predictable environmental dynamics and b. severe communication constraints. We will describe recent results on the design of control algorithms for such a robot team engaged in an information gathering task.


Mark Campbell - Distributed Sensor Fusion and its Impact on Resilient Autonomy

Resilient autonomy can be characterized by reliable and improving performance over time, in the presence of uncertain data and uncertain communications. For multi-agent teams, resilient autonomy relies fundamentally on strong coupling between distributed sensor fusion/perception and distributed planning/control. This talk will explore distributed sensor fusion concepts, and their implications on distributed planning and control - key enablers for resilient autonomy. First, distributed sensor fusion theory will be presented as a function of several key components, including scalability, process complexity, sensor dynamics/uncertainty/heterogeneity, and network topology/uncertainty. Importantly, while many of these individual problems can be addressed, it is the combination where problems arise. A key goal of the distributed sensor fusion methodology is to create individual and collective knowledge and confidence, for subsequent planning/control/decision making. The rest of the talk will present implications of the distributed sensor fusion issues on planning and control.


Alejandro Ribeiro - Online Learning for Characterizing Unknown Environments in Robotic Vehicle Models

In pursuit of increasing the operational tempo of a robotics platform in unknown domains, we consider the problem of predicting the distribution of state-estimation error due to poorly-modeled platform dynamics as well as environmental effects. Such predictions are a critical component of any modern control approach that utilizes uncertainty information to provide robustness in control design. We use an online learning algorithm based on matrix factorization techniques to fit a statistical model of error that provides enough expressive power to enable prediction directly from motion control signals and low-level visual features. Moreover, we empirically demonstrate that this technique compares favorably to predictors that do not incorporate this information.


Geoffrey Hollinger - Information Gathering with Multi-Robot Teams: Towards Resilient Shared Autonomy with Humans

Typically when teams of robots are tasked with gathering information (e.g., in urban search and rescue, environmental monitoring, and aerial surveillance scenarios), human operators must oversee almost every aspect of the operation to ensure completion of the task. Strict human oversight not only makes such deployments expensive and time consuming but also makes some tasks impossible due to the requirement for heavy cognitive loads or superhuman reaction times. Scaling up the team size to 100s or 1000s of heterogeneous robots makes these issues even more apparent, since humans cannot feasibly control the low-level behaviors of such large teams. To mitigate the limitations described above, we propose developing shared autonomy architectures, where the human operator provides high-level goals, and the robotic team determines its own low-level actions. In this talk, I will show how a general framework that unifies information theoretic optimization and physical motion planning makes shared autonomy tractable in exploration, monitoring, and mapping domains. The framework leverages techniques from reinforcement learning and sampling-based motion planning to provide scalable solutions in a diverse set of applications, such as urban target search and marine monitoring.


Group 2: Security and Privacy

Xenofon Koutsoukos - Resilience in Networked Dynamic Systems using Trusted Nodes

Local interaction rules for consensus and synchronization are vital for many applications in distributed monitoring and control of robot networks. However, typical methods assume all nodes (or agents) in the networked system cooperate. Our work considers local interaction rules for resilient consensus and synchronization in the presence of adversary nodes. Typical algorithms require networks to have a high connectivity to overrule the effects of adversaries, or require access to non-local information. Such algorithms increase the attack surface and they may compromise the security of robot networks. We propose a scheme for resilient distributed consensus using a set of trusted nodes within a network. A subset of nodes, which are more secured against the attacks, constitute a set of trusted nodes. Using this notion, we provide an alternative way to guarantee distributed consensus despite any number of adversarial attacks, even in sparse networks. It is shown that the network becomes resilient against any number of attacks whenever the set of trusted nodes form a connected dominating set within the network. Further, we show that the overall network connectivity is significantly improved by making a very small subset of nodes (edges) trusted, even if the original network is sparse. Moreover, by controlling the number of trusted nodes, we can obtain any desired level of connectivity.


George Pappas - Security and Privacy of (Multi)-Robot Systems

The interaction between information technology and robots makes robots vulnerable to malicious attacks beyond the standard cyber attacks. This has motivated the need for attack-resilient state estimation. Yet, the existing state-estimators are based on the non-realistic assumption that the exact system model is known. Consequently, in this talk we present a method for state estimation in presence of attacks, for systems with noise and modeling errors. When the the estimated states are used by a state-based feedback controller, we show that the attacker cannot destabilize the system by exploiting the difference between the model used for the state estimation and the real physical dynamics of the system. This enables mapping control performance requirements into real-time (i.e., timing related) specifications imposed on the underlying platform. Finally, we illustrate and experimentally evaluate this approach on an unmanned ground vehicle case-study. We conclude by presenting some problems for future research in multi-robot security and privacy.


Amanda Prorok - Quantifying Privacy for Tightly Coupled Heterogeneous Robot Swarms

We are interested in securing the operation of robot swarms composed of heterogeneous agents. Since any given robot type plays a role that may be critical in guaranteeing continuous and failure-free operation of the system, it is beneficial to hide information that reveals the individual robot types and, thus, their roles. We propose a method that quantifies how easy it is for an adversary to identify the type of any of the robots, based on an outside observation of the system’s behavior. We draw from the theory of differential privacy, and develop an analytical model of the information leakage that depends on parameters of the swarm behavior. We look to further this line of work by developing active privacy mechanisms that are able control the amount of information leaked, while maintaining the overall performance of the underlying swarm system.


Group 3: Collaboration & Coordination in Heterogeneous Teams

Dylan Shell - On the interplay of explicit and implicit coordination in multi-robot systems

In this talk I will describe two pieces of research in which we used traditional multi-robot task-allocation mechanisms to coordinate robots -- but in which the robots had to interact with implicitly coordinating elements of the robotic system too. This topic is relevant to the workshop because it can be used to coordinate robot systems where individuals have quite diverse capabilities. I plan to share ideas about how one set of robots can interact with a set of others by: (1) modeling collective behavior; (2) using these models to generate geometric primitives; (3) realizing those primitives within the environment.


Pratap Tokekar - Algorithms for Heterogeneous Robotic Data Collection

A connected network of robots, sensors, and smart devices has the potential to solve grand challenges in domains such as agronomy, oceanography, and emergency response. Robots will form the "physical" layer of this network and collect data from hard to reach places at unprecedented spatio-temporal scales. Heterogeneity in sensing and mobility in these robot teams is critical in order for us to effectively collect data from diverse, unstructured, natural environments. In this talk, I will present our recent work on devising efficient algorithms for data collection with heterogeneous robot teams. We will focus on the routing and coordination of aerial and ground robots acting as robotic cameras. I will present approximation algorithms for multi-robot teams to map known and unknown environments in the least amount of time. I will discuss our ongoing work on a heterogeneous version of the Traveling Salesman Problem used to assign goals for the aerial and ground robots. Finally, I will describe our new work on informative path planning for coordinated sampling in marine environments with aerial robots and robotic boats.


Lynne Parker - Collaboration in communications-constrained heterogeneous teams

In many types of heterogeneous robot-robot and human-robot collaborations, high-bandwidth communications are not possible or desirable. It may thus be impractical to communicate sufficient information between team members to achieve effective collaborations in real time. Further, the heterogeneity and complexity of the team members often makes it difficult to build detailed models that can predict the behavior of team members across a wide variety of collaborative states. Thus, methods of implicitly interpreting and understanding the state, intent, and actions of heterogeneous team members are required, in order to generate meaningful collaborations. This talk discusses a variety of heterogeneous teammate interactions, and suggests challenges and methods for the implicit understanding of human and/or robot activities in collaborative teams.


Rafael Fierro - Exploiting Heterogeneity in Robotic Networks

We exploit heterogeneity in robotic systems by taking advantage of complementary communication technologies, distinct sensing functionalities and different motion dynamics present in the network. Complex missions are accomplished by assigning specialized sub-tasks to specific members of the robotic system. An adequate selection of the team members and an effective coordination are some of the challenges to fully exploit the unique capabilities that these types of systems can offer. Motivated by real world applications, we focus on a multi-robotic network made by aerial and ground agents which has the potential to provide critical support to human teams in dynamic, partially known environments. For instance, aerial robotic relays are capable of transporting small ground mobile sensors and also of expanding the communication capabilities of the whole system. More specifically, maintaining reliable communications on heterogeneous robotic networks is fundamentally important, especially for cooperative autonomy. Radio communications are the common method which allows the robots to operate wirelessly. However, this technology has some limitations that affect the capacity of the robotic network. Optical wireless communication has been proposed as an ideal complement to mitigate some of the weaknesses of radio frequency systems. Therefore, we are motivated to study designs for a hybrid RF/Optical wireless link between aerial platforms and mobile ground robots. In this work, our intention is to combine radio frequency communications and optical data transfer to have robust connectivity.


Volkan Isler - Heterogeneity and Sensor Planning

Robots with complementary sensing and mobility capabilities can enable spatial and temporal coverage of dynamic phenomena at unprecedented scales. In my talk, I will give examples of this capability from our ongoing work on environmental and agricultural monitoring. I will conclude with a broad set of research challenges.


Group 4: Resilient Control

Calin Belta - Resilient Formal Synthesis

In control applications, one popular approach to deal with changing environments is model predictive control (MPC), in which a cost is iteratively optimized over a time horizon. Robust versions of MPC have been shown to provide (sub)optimal solutions for cases in which the system parameters are uncertain. In this talk, I will show results on robust MPC under correctness requirements formulated as timed temporal logic constraints. I will then discuss the case when the process is not known, and show how the particular semantics of the logic that we use allows to map the control problem to a reinforcement learning problem. Robotics case studies will be used for illustrations, both in single-agent and multi-agent scenarios.


Seth Hutchinson - Robust Distributed Control Policies for Multi-Robot Systems

In this talk, I will describe our recent progress in developing fault-tolerant distributed control policies for multi-robot systems. We consider two problems: rendezvous and coverage. For the former, the goal is to bring all robots to a common location, while for the latter the goal is to deploy robots to achieve optimal coverage of an environment. Each robot has a limited sensing range, and is able to directly estimate the state of only those robots within that sensing range, which induces a network topology for the multi-robot system. We assume that it is not possible for the fault-free robots to identify the faulty robots. For each problem, we provide an efficient computational framework and analysis of algorithms, all of which converge in the face of faulty robots under a few assumptions on the network topology and sensing abilities.


Shreyas Sundaram - Scalable algorithms for resilient distributed coordination in large-scale networks

In this talk, we describe a class of scalable and lightweight distributed algorithms for overcoming malicious behavior by nodes in large-scale networks. Starting with the canonical problem of distributed consensus, we consider a class of algorithms where at each time-step, each node in the network disregards its most extreme neighbors and updates its value as a weighted average of the remaining values. We show that traditional graph metrics such as connectivity and node degree are not sufficient to analyze this class of dynamics. Instead, we introduce a quantifiable graph-theoretic property termed "robustness," and show that consensus is guaranteed under the above dynamics despite the actions of a potentially large set of malicious nodes, provided that the graph is sufficiently robust. We also show that common mathematical models for large-scale networks inherently possess such robustness properties. We conclude by describing how the concept of "local-filtering" can be applied to address the problems of distributed optimization and distributed state estimation in the presence of adversarial nodes.


Sanjit Seshia - Diversity and Resilience through Control Improvisation

One approach to handling an unexpected situation is to improvise. But what does it mean, precisely, for robots to improvise? In this talk, I will describe control improvisation, an attempt at formalizing the notion of improvisation for use in the design, verification, and deployment of cyber-physical systems. I will sketch both directions for theoretical work and for applications with a focus on better understanding the notions of diversity and resilience for multi-robot systems.


Group 5: Multi-Robot Planning & Control

Maxim Likhachev - Search-based Planning by Decomposition for Teams of Heterogeneous Robots

Presenting a planning problem as a graph search and then searching the graph for a close-to-optimal solution is a popular planning approach for single-robot systems. For multi-robot systems though, graph search-based planning approaches are less popular due to the computational complexity of searching the graph that represents the joint state-space. At the same time, graph representation provides multiple benefits, in particular with respect to handling well complex constraints, arbitrary cost functions and heterogeneity of robots in the team. In this talk, I will briefly overview our research on developing search-based planning techniques that do provide real-time performance in coordination of a team of heterogeneous robots. The main insight these techniques rely on is that many real world problems are not arbitrary complex and instead can often be decomposed into a series of easier-to-solve graph searches that converge to solutions with provable guarantees on quality. I will overview several of these techniques and their applications to reconnaissance and navigation with collaborative localization.


Stefano Carpin - Risk aware multi-robot planning

When multi robot systems operate in complex, dynamic scenarios, failure of some or components is an unavoidable possibility. Being able to quantitatively assess these failure risks is an essential step for design and analysis purposes. In this talk we will review our recent work on risk aware planning in multi robot systems and will present a data-driven framework based on constrained Markov Decision Processes for multi-robot planning.


Marco Pavone - Models, Algorithms, and Evaluation for Autonomous Mobility-On-Demand Systems

In this talk I will discuss the operational and economic aspects of autonomous mobility-on-demand (AMoD) systems, a rapidly emerging mode of personal transportation wherein robotic, self-driving vehicles transport customers in a given environment. Specifically, I will discuss AMoD systems along three dimensions: (1) modeling -- analytical models capable of capturing the salient dynamic and stochastic features of customer demand, (2) control -- coordination algorithms for the vehicles aimed at throughput maximization, and (3) economic -- fleet sizing and financial analyses for case studies of New York City and Singapore. I will conclude the talk by presenting a number of directions for future research.


Spring Berman - A Control and Estimation Framework for Heterogeneous Robotic Swarms with Stochastic Behaviors Models

We are developing a rigorous control and estimation framework for robotic swarms that are deployed into unknown environments. This framework will apply to heterogeneous swarms of robots that have diverse capabilities and dynamically changing roles. A heterogeneous swarm may be comprised of numerous low-cost, resource-constrained platforms, which are restricted to local information obtained from random encounters, and a few sophisticated platforms with global localization and greater computational capabilities. Our framework will enable these swarms to operate largely autonomously, with user input consisting only of high-level directives that map to a small set of robot parameters. We use stochastic and deterministic models from chemical kinetics and fluid dynamics to describe the robots’ roles, task transitions, spatiotemporal distributions, and manipulation dynamics at both the microscopic (individual) and macroscopic (population) levels. In this talk, I will describe our work on various aspects of the framework, including strategies for environmental mapping, scalar field estimation, boundary coverage, cooperative manipulation, formation control, and herding. We will validate our techniques on a heterogeneous multi-robot testbed that includes a swarm of small ground robots and several aerial vehicles.


Magnus Egerstedt - Encoding Heterogeneity Through Constraints In Multi-Robot Teams

When designing coordinated control algorithms for achieving team-level objectives, safety and networking constraints are either explicitly taken into account already at the design stage, thereby making the design task significantly more complex, or added at a later stage. In this talk, we show how one can incorporate these types of lower-level considerations as constraints using Control Barrier Functions in a provably, minimally invasive manner, i.e., in such a way that the team-level objectives are respected as much as possible. Moreover, one can encode heterogeneous capabilities and objectives through these types of constraints as well, resulting in not only a richer set of behaviors but also in improved performance as compared to homogeneous formulations. In fact, in community ecology, richness of behavior is largely derived from constraints, such as scarcity/abundance of food sources or predators, and the talk will make formal connections to these ecological principles as well as deploy the developed framework on a team of mobile robots.


Nora Ayanian - Using Online Games to Inspire Distributed Multirobot Controllers

Humans are quite good at coordination, due to our computational, communication, and sensing capabilities, but also due to diversity. In human groups, order emerges because some humans are better at some tasks than others. This may lead one to ask: Can we get teams of robots to coordinate like humans? But robots have considerably limited computation, communication, motion, and sensing capabilities. Instead, then, the question we might want to ask is: Can humans coordinate effectively under robot-like conditions? We have attempted to answer this question by conducting investigations on the ability of humans to solve challenging collective coordination tasks in a distributed fashion with limited perception and communication capabilities similar to those of a simple ground robot. In this talk, I will motivate this research and highlight some interesting observed behaviors, including distinct traits in the human behaviors. Preliminary analysis suggests a set of diverse, stochastic controllers learned from human behavior may help solve some challenging multirobot tasks in a distributed way.


Ronald Arkin - The Value of Slowness in Robotic Systems

Robot teaming is a well-studied area, but little research to date has been conducted on the fundamental benefits of heterogeneous teams and virtually none on temporal heterogeneity, where timescales of the various platforms are radically different. This talk explores this aspect of robot ecosystems consisting of fast and slow robots (SlowBots) working together, specifically considering the aspect of slowness. Slowness in robotic systems is a quality that is typically undervalued. It is our contention that as slowness has utility in animal behavior in certain species that it can also provide useful qualities for robotic implementations in appropriate circumstances. In particular we study mammalian behavior as evidenced in the tree sloth and slow Loris as the basis for the behaviors of a robot capable of residing in an arboreal ecological niche, as might be found in certain jungle surveillance applications or certain agricultural tasks. We also describe the entrainment of such an agent to the circadian rhythms present in nature that will benefit its long-term persistence in the environment.