Monitoring a dynamic environment with robot teams requires continuously solving the multi-robot task allocation (MRTA) problem in response to environmental changes. The adaptive assignment of robots to different parts of the workspace as the environment changes makes this a Single-Task robots, Multi-Robot tasks, and Time-extended Assignment (ST-MR-TA) problem. Solutions to this problem can be classified as either macroscopic, where the assignment is obtained using a mean-field model of the team dynamics, or microscopic, where the assignment is posed as a resource allocation problem. While macroscopic techniques are scalable with team sizes and number of tasks, they lack expressiveness. On the other hand, microscopic techniques are more expressive, but often require expensive replanning when environmental conditions change. In this work, we propose an alternative macroscopic formulation to the ST-MR-TA problem that results in a time-varying task assignment that can correspond to a changing environment. Our analysis of the macroscopic model uncovers parameter regimes where time-varying populations exist and further investigates the break down when the system is not well-mixed, i.e., when team sizes are small. Simulation validation shows that our proposed feedback control is necessary for small teams.
We tackle planning under uncertainty when multiple robots must proactively plan perception and communication acts, and decide whether the cost needed to obtain a state estimate is justified by the benefit of the information obtained.The approach is suitable when observations are costly but, when they do occur, are of high quality and recover the system’s joint state, either alone or along with communication. Such cases allow one to sidestep the construction of the full joint belief space, a well known source of intractability in planning. Formulating the problem as a class of Markov decision processes to be solved over joint states and structured to allow decentralized execution, we give a suitable Bellman recurrence using macro-actions.We solve for policies for the individual robots, providing a simulation case study and reporting on a physical robot implementation. Based on our experience with hardware, we examine some non-idealities identified in practice, proffering suitable enhancements to the basic model.
In this paper, we examine the adaptability of a multi-robot coordination system in which distributed autonomous robots respond to unexpected situations based on functional expressions using large language models (LLMs). In recent years, there has been growing interest in systems where multiple autonomous distributed construction machines (robots) collaborate and adaptively perform tasks in open and unknown environments, including disaster sites. In open environments, unforeseen events that cannot be predicted in advance may occur, and it is challenging to address these events solely with existing model-based approaches. In this paper, we leverage the high environment comprehension capabilities of foundation models to understand unforeseen situations and develop a system that enables adaptive coordinated actions by flexibly integrating the functions of multiple robots using LLMs. Additionally, we account for the embodiment (interactions between the robots and their environment). Our experiments demonstrate that the designed system is capable of adaptively responding to three types of unforeseen situations, including path obstructions caused by either an obstacle or a robot. In cases of path obstructions caused by obstacles of varying weights, the system can exhibit appropriate obstacle removal behaviors by reflecting the torque capability of the robot as one aspect of its embodiment.
This paper presents GTernal, an open-source robot designed for multi-robot experimentation with a focus on reducing the overhead labor required to setup and conclude multi-robot experiments. This paper discusses the robot’s hardware design choices and the supporting algorithms that together enable the automation of the time intensive initialization, clean up, and charging routines associated with multi-robot experimentation. The efficacy of the presented design and algorithms are validated by the continuous execution of 453 experiments over a three day period on the Robotarium testbed.
The integration of simple robots into real-world scenarios necessitates their capacity to collaborate within small groups while effectively managing system faults. In this work we introduce the problem of robust matching maximization, in which a group of simple distributed decision-making robots should match in pairs, despite the activity of faulty robots within the group. These faulty robots, indistinguishable and prone to disruptive behavior, may prevent the nonfaulty robots from successfully meeting. We therefore devise distributed algorithms aiming at maximizing the number of pairs assembled by the nonfaulty robots. By establishing both lower and upper bounds on the achievable pairings, we explore the algorithms theoretically, and offer a rigorous empirical analysis, testing their robustness to various types of failures, robot characteristics, and fault intensity.
A swarm robotic system, consisting of numerous distributed autonomous robots, has been increasingly envisioned for applications in coordinated object removal tasks, such as clearing debris and fallen trees in unknown and complex environments typical of disaster sites. In these environments, the system should be robust, flexible, and scalable, and able to function even when communication between robots is interrupted. Although some previous studies have demonstrated flexible cooperative behaviors using the centralized approach, there is still no autonomous distributed swarm robotic system that takes these considerations into account. This paper aims to develop a coordinated system of distributed autonomous swarm robots with robustness, flexibility, and scalability by applying the Transformer encoder to swarm systems. The focus of this study is on coordinated object removal tasks. Using the TurtleBot3 Burger robot model, we train a coordinated object removal behavior that operates without any communication between robots, leveraging reinforcement learning. We demonstrate that this behavior can be applied to scenarios involving varying numbers and sizes of objects, as well as different quantities of robots, without the need for retraining. We further explore the effect of noise on sensor data and motor performance. Additionally, we demonstrate the application of our system in a rescue scenario within an unknown environment, where it coordinates with another autonomous mobile robot to enhance effectiveness.
In this study, we propose a synchronized motion-switching method for a robotic swarm comprising autonomous distributed mobile robots having only local communication ability. Assuming that a robot is equipped with a phase oscillator, we regarded the robotic swarm as a coupled oscillator system (COS). Furthermore, we propose wave-type local interactions among oscillators, which generate stationary-phase waves over a network of COS. Because a phase wave depends on the graph structure of the COS, a robot can determine the global changes in a swarm using spectrum analysis of its phase. The proposed system was applied to a robotic swarm to make synchronized decisions to change the direction of movement. Without a synchronization system, a robotic swarm can be split into pieces. Contrastingly, the proposed system can maintain a unified behavior even if a robot has no global communication or leader. We verified the feasibility of local interactions by building real robots, and examined the effectiveness of the simulations.
Using a swarm of robots introduces unique challenges in how to simultaneously operate a large number of robots. Testing large-scale swarm algorithms with hardware-in-the-loop is difficult without local, manual intervention. Building tools to handle operations for a large swarm such as charging, positioning, powering, and programming allows developers to use the swarm in an effective manner. Furthermore, building upon these operations to allow a swarm to function without direct human involvement enables it to be usable by local and remote programmers alike. These remotely accessible swarm test beds generally require capabilities such as, automatically executing algorithms, gathering data, and charging the robots as needed. In order to constitute a scalable swarm, it is also crucial that the robots are cost-effective, easy to build, and straightforward to operate. This paper introduces the Coachbot Swarm Testbed, a 100-robot swarm, details how it achieves these crucial operations to function without human intervention, and explains how programmers, both local and remote, can utilize the swarm to test their algorithms.
By exchanging information, multiple robots can compensate for the limitations in their individual perceptions of their surroundings. Our previous study proposed a two-step cooperative inference model using foundation models to assess diverse robot actions based on the information exchanged in natural language with other robots and humans. The inference model classifies executed actions into three categories—accomplished, not accomplished, and unclear—using the fuzzy inference system and Dempster-Shafer theory. However, the inference model’s reliance on a single rule in fuzzy inference raises concerns about the accuracy of its action classification. Additionally, evaluating the model’s reliability, validity, and efficiency remains unsatisfactory. In this paper, we first enhance the inference model to generate multiple fuzzy rules and perform a comparative evaluation of reliability, validity, and efficiency using a dataset from a real-world task. Then, we discuss the evaluation results of the first and second steps, the effect of introducing the unclear category in preventing incorrect judgments, and the effect of threshold adjustments in the second step on the inference results.
Cooperative decentralized learning relies on direct information exchange between communicating agents, each with access to locally available datasets. The goal is to agree on model parameters that are optimal over all data. However, sharing parameters with untrustworthy neighbors can incur privacy risks by leaking exploitable information. To enable trustworthy cooperative learning, we propose a framework that embeds differential privacy into decentralized deep learning and secures each agent’s local dataset during and after cooperative training. We prove convergence guarantees for algorithms derived from this framework and demonstrate its practical utility when applied to subgradient and ADMM decentralized approaches, finding accuracies approaching the centralized baseline while ensuring individual data samples are resilient to inference attacks. Furthermore, we study the relationships between accuracy, privacy budget, and networks’ graph properties on collaborative classification tasks, discovering a useful invariance to the communication graph structure beyond a threshold.
In this paper we present the design and the control of a swarm of bimodal particles that switch their geometric shape between two modes. The designed particle has the shape of a dodecahedron and is 3D printed using layers of thermoplastic polyurethane and polylactic acid materials. Using the design parameters the particles are such that they react to an external stimuli of temperature to induce the switching between the two modes: open and closed. The motion model of the swarm was identified to be that of a Brownian particle for the open mode and the noisy unicycle for the closed mode corresponding to different parameters for each mode. To effectively model the noise parameters affecting the motion of the particles we used experimental validation. Using the experimental validation we aimed to determine switching control of the swarm based on the Motility-Induced Phase Separation (MIPS) index that quantifies the aggregation of the swarm. To characterize the switching of the particles we conducted a simulation in MATLAB using the noise parameters identified and determined aggregation of the swarms in both modes using the MIPS index. Our simulations demonstrate that using the noise parameters identified, desired swarm aggregation can be achieved with simple robots that are capable of changing their geometric shape. We draw attention to how simple hardware design of a single agent can achieve aggregations enabling sensing-related tasks to be completed.
Modular self-reconfigurable robots can be used to build objects either by filling the space entirely with robots (densely packed) or by constructing a hollow internal structure (scaffolding) and coating the outside. Scaffolding offers many advantages for building an object versus a densely packed representation. However, the different approaches have never been compared. In this article, we define and evaluate metrics for comparing scaffolding algorithms and we test them on three existing scaffolding algorithms.
Reconfigurable robot swarms exhibit the ability to perform collaborative tasks, change morphologies, and traverse challenging terrains beyond the capability of a single robot. On mobile robot platforms, it is challenging to incorporate both individual mobility, and coupling and reconfiguration capabilities. In this paper, we present two mechanisms to enable obstacle ascent and descent behaviors to enhance the locomotion capabilities of reconfigurable swarm robots, based on the PuzzleBot swarm. These mechanisms, simple 1-DOF arms attached to the front of the robots, leverage the swarm’s flexible, passive coupling to reduce energy consumption when performing these added locomotion tasks. They also only require a single added mechanism to perform these tasks, using the passive coupling to continue behaviors once initiated by the mechanisms. We present the design and modeling of these mechanisms, as well as the various behaviors enabled by the added degree of freedom. We show the results of mechanism characterization and hardware experiments that evaluate the gap crossing, obstacle descent, climbing, and lifting behaviors of the reconfigurable robot swarm. Compared with the previous PuzzleBots, we demonstrate significant improvement in the above-mentioned tasks with controlled experiments. Finally, we show that the robots can perform in real-world scenarios outside of a lab environment. Video available at https://youtu.be/-VvjX-nyTUE .
We present a novel coordination scheme for online multi-robot coverage with a team of resource-constrained aerial vehicles relying on the PH-tree data structure. We propose a hybrid system architecture where coverage tasks are assigned to robots in a centralized manner, and individual robots plan their paths cooperatively in a distributed fashion. In this way, robots can focus only on local path planning tasks, while a ground station ensures coordination among the team. We show the applicability of our method to various objects and quantitatively evaluate the performance of our solution for various robot team sizes.
Multi-target tracking in various security, surveillance, and reconnaissance tasks involves a large number of moving targets moving around in large workspaces motivating the use of mobile, distributed, and cooperative robot swarms. Incorporating human operators in these multi-robot multi-target tracking tasks can make the robot swarms more adaptive, increase robustness and scalability, and hence improve the overall tracking performance. However, collaboration in these mixed human-robot teams (MHRTs) raises concerns about how to enable communication between the human operator and the robot swarm, and how to design robot planning and control algorithms that efficiently leverage information from the human operator. This paper presents a cooperative tracking framework for multi-robot multi-target tracking. A novel mixed-initiative cooperation strategy to facilitate communication between the human and robots is presented. A new tracking utility function is developed to incorporate human inputs and enable a distributed robot network equipped with directional field-of-view (FOV) sensors to track multiple moving targets simultaneously in obstacle-populated environments. An online target state estimation framework is developed to leverage learning-based perception algorithms for target detection, classification, and estimation for tracking. This target state estimation framework is used to test the cooperative tracking framework with a team of ground robots equipped with RGBD cameras. Using simulation studies and physical experiments it is shown that this framework can provide robust performance in the presence of uncertainties such as state estimation errors and intruders and consistently outperform homogeneous robot teams.
In this paper, we propose a fully distributed method for scheduling the movements of modular robots in a set while maintaining its balance. We explore self-reconfigurable modular robots, which are robots that can change their shape and behavior on their own. We focus on the important problem of verifying their stability and keeping these robots stable while they change shape. Our approach involves developing an algorithm that helps the robots stay balanced in real-time, even when they are changing shape. At the heart of this algorithm is the mechanical notion of the “support polygon” and the concept of “Mobile Object ” in message data, which enables a fast stability verification. We test this algorithm using a programmable matter simulator called VisibleSim, we show that it works well for validating a current static situation and for predicting the status after one or more movements. Additionally, we introduce the concept of dimensional reduction from 3D to 2D, simplifying stability analysis while maintaining accuracy. By addressing stability challenges, our research aims to make self-reconfigurable modular robots more useful in various applications.
Robotic swarms often cite scalability as a benefit of their deployment. To realize the full extent of this benefit, swarm behavior algorithms have to be designed so that performance is not inhibited at scale. In this work, we examine a set of persistent and adaptable swarm shape formation algorithms that are limited by onboard memory when executed at scale. We improve these algorithms by using robot communication and collective knowledge to allow robots to form persistent and adaptive shapes with a fixed memory dependence. We demonstrate the improved algorithms in both simulation and on a swarm of mobile robots.
This paper explores variants of ReactiveBuild, an algorithm that enables 3D, free-form, and environment-adaptive robot self-assembly. The most successful variant introduces a population of “arrogant” agents that ignore low-priority recruitment signals. This variant improved the simulated self-assembly of structures while also reducing the number of steps required for robots to create these structures. These results suggest that swarm heterogeneity can enhance the performance of self-assembly algorithms. Additionally, the paper discusses several variants that did not meaningfully improve algorithm performance but nonetheless provide interesting and useful lessons for robot self-assembly researchers. These variants were the introduction of hysteresis, varying the sensitivity of robots to force measurements with time, randomly moving in an incorrect direction, and probabilistic stopping. The limited impact to algorithm performance indicates that the ReactiveBuild algorithm is robust against robot hardware issues such as force sensor errors, environmental perception issues, and inconsistent communications, suggesting that the algorithm is well-suited to deployment on real robot hardware.
In the research community, there’s a keen interest in the challenges faced when drones are deployed for tasks like surveillance, environmental monitoring, and search-and-rescue operations. Fundamental behaviors such as coverage control and area monitoring are crucial for dynamic missions. This paper introduces the results of a control algorithm for monitoring points of interest within the context of drone deployment. The goal is not only to present a novel approach based on Voronoi diagrams and Control Barrier Functions but also to demonstrate the algorithm’s ability to maintain formation and complete tasks in scenarios where the drone swarm encounters obstacles in an environment whose map is unknown a priori. To validate the effectiveness of our approach, we conducted an experimental study featuring a comprehensive suite of simulations, where the working conditions and potential environmental obstacles were not known in advance.
Collective motion is a widespread natural phenomenon in which individuals move in an ordered fashion, without centralized control. Such movement is often considered to be robust to individual failures, but common models of collective motion fail to display this trait. Inspired by observations of locusts, we introduce intermittent pauses in individual motion, where pause duration is tied to perception of neighbors’ movements. Realistic physics-based simulations of swarm robots demonstrate that this leads to significant improvements in the resilience of the swarm, without sacrificing its functionality or coherence.
