Paper deep dive
A Prototyping Framework for Distributed Control of Multi-Robot Systems
Junaid Ahmed Memon, Allan Andre Do Nascimento, Kostas Margellos, Antonis Papachristodoulou
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 93%
Last extracted: 7/8/2026, 3:39:32 PM
Summary
This paper introduces a unified prototyping framework for distributed control of multi-robot systems that bridges theoretical distributed optimization algorithms with practical testing. Utilizing the Single Program, Multiple Data (SPMD) paradigm, the framework emulates decentralized control on a single multi-core workstation, supporting three fidelity levels: simple dynamical models, high-fidelity digital twins, and physical hardware (Crazyflie quadcopters). The approach is demonstrated on a four-UAV position-swapping task using a non-cooperative game-theoretic algorithm, proving it to be a low-cost, modular, and accessible solution for validating distributed control strategies.
Entities (6)
Relation Signals (5)
Distributed Control Prototyping Framework → uses → SPMD Paradigm
confidence 95% · Using the Single Program, Multiple Data (SPMD) paradigm, the framework emulates distributed control on a single computer, with each core running the same algorithm using local states and neighbour-to-neighbour communication.
Distributed Control Prototyping Framework → validateson → Crazyflie Quadcopters
confidence 93% · Computational time and trajectory data are compared across the supported dynamics levels: a point-mass model, a high-fidelity quadrotor model, and an experimental hardware testbed using Crazyflie quadcopters.
Crazyflie Quadcopters → utilizes → Lighthouse-V2 System
confidence 92% · Localization is provided by the Lighthouse-V2 infrared motion-tracking system, where onboard photodiodes detect laser sweeps from multiple base stations to triangulate the UAV pose with approximately 1 cm accuracy
Distributed Control Prototyping Framework → implements → Non-cooperative Game-theoretic Algorithm
confidence 90% · We demonstrate the framework on a four-quadrotor position-swapping task using a non-cooperative game-theoretic distributed algorithm.
Distributed Control Prototyping Framework → developedusing → MATLAB Parallel Computing Toolbox
confidence 88% · The framework was implemented using the MATLAB® Parallel Computing Toolbox, where agents execute in parallel within an SPMD block and each computational engine corresponds to a worker.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:This paper presents a prototyping framework for distributed control of multi-robot systems, aimed at bridging theory and practical testing of distributed optimization algorithms. Using the Single Program, Multiple Data (SPMD) paradigm, the framework emulates distributed control on a single computer, with each core running the same algorithm using local states and neighbour-to-neighbour communication. We demonstrate the framework on a four-quadrotor position-swapping task using a non-cooperative game-theoretic distributed algorithm. Computational time and trajectory data are compared across the supported dynamics levels: a point-mass model, a high-fidelity quadrotor model, and an experimental hardware testbed using Crazyflie quadcopters. The results show that the framework provides a low-cost and accessible approach for validating distributed algorithms.
Tags
Links
- Source: https://arxiv.org/abs/2605.15049v1
- Canonical: https://arxiv.org/abs/2605.15049v1
Trouble viewing inline? Open PDF directly →
Full Text
32,426 characters extracted from source content.
Expand or collapse full text
A Prototyping Framework for Distributed Control of Multi-Robot Systems Junaid Ahmed Memon Allan Andre Do Nascimento Kostas Margellos Antonis Papachristodoulou Department of Engineering Science, University of Oxford, UK (e-mail: junaid.memon, allan.adn, kostas.margellos, antonis@eng.ox.ac.uk) Abstract This paper presents a prototyping framework for distributed control of multi-robot systems, aimed at bridging theory and practical testing of distributed optimization algorithms. Using the Single Program, Multiple Data (SPMD) paradigm, the framework emulates distributed control on a single computer, with each core running the same algorithm using local states and neighbour-to-neighbour communication. We demonstrate the framework on a four-quadrotor position-swapping task using a non-cooperative game-theoretic distributed algorithm. Computational time and trajectory data are compared across the supported dynamics levels: a point-mass model, a high-fidelity quadrotor model, and an experimental hardware testbed using Crazyflie quadcopters. The results show that the framework provides a low-cost and accessible approach for validating distributed algorithms. keywords: Distributed Control, Multi-robot Systems, Digital Twins, Parallel Optimization. †thanks: JAM acknowledges support from the Oxford Pakistan Programme. AAdN, AP, and KM acknowledge support from MathWorks. AP was additionally supported by EPSRC grants EP/X017982/1, EP/Y014073/1, and UKRI2108. For Open Access, the authors have applied a C BY licence to any Author Accepted Manuscript arising from this submission. 1 Introduction Multi-robot systems are increasingly used in demanding tasks, such as environmental monitoring, inspection, search and rescue where scalability, performance, and autonomy are needed. Although distributed optimisation and coordination algorithms enable scalable decision making, their physical validation remains challenging. Bridging simulation and hardware gap therefore requires modular, affordable, and reproducible testbeds that capture realistic sensing, communication, and computation constraints Mokhtarian et al. (2024). Testing distributed algorithms on robot fleets faces two main challenges. First, although distributed architectures improve scalability, fault tolerance, and reliability over centralised ones (Jamshidpey et al., 2025), switching between architectures or adapting code across them is often difficult. Second, simple models, digital twins, and hardware controllers are typically developed separately, increasing effort and slowing prototyping. A unified framework that preserves the robot-fleet structure across all testing stages is therefore essential for reliable, safe, and stable deployment (Shakeri et al., 2019), while accelerating the design, build and test cycle. 1.1 Related Work Distributed optimisation algorithms are well studied for multi-agent systems Notarstefano et al. (2019); Yang et al. (2019). Despite advances ranging from ADMM-based formation control (Stomberg et al., 2023) to decentralised sequential quadratic programming for non-linear MPC (Stomberg et al., 2025), real-time hardware implementation remains difficult due to iterative communication, timing constraints, and heterogeneous agent dynamics. Although simulation studies are abundant, only few support seamless hardware transition Shorinwa et al. (2024). For experimental validation, several multi-robot testbeds have been developed. Scaled transport platforms such as the University of Delaware scaled smart city (Stager et al., 2018) demonstrate cooperative vehicle coordination in structured environments. Swarm platforms such as Kilobot (Rubenstein et al., 2014) and Robotarium (Pickem et al., 2017; Wilson et al., 2020) provide low-cost distributed coordination environments, but are typically limited to a single class of robot dynamics, e.g. differential-drive ground robots, and rely on localization setups that are difficult or expensive to replicate. Duckietown (Paull et al., 2017) and the Cambridge RoboMaster testbeds (Blumenkamp et al., 2024) extend capability, but remain tied to specific hardware or sensing modalities, making adaptation to new robot classes or distributed optimisation routines non-trivial. For multi-UAV platforms, the systems in Deng et al. (2020) and the distributed hardware implementations in King (2004) are particularly notable. The range of available platforms reflects growing interest in experimental validation of distributed systems. However, prior work often requires dedicated hardware clusters, complex networking, or high-cost setups, limiting accessibility for small-scale academic and teaching environments. Moreover, to the best of our knowledge, none provides integrated simulation and hardware testing within a single framework that eliminates multiple tools and reduces testing time for distributed multi-robot algorithms. 1.2 Contributions To address these shortcomings, this work proposes a unified parallel-computing framework for rapid prototyping of distributed control algorithms across three fidelity levels: simplified models, high-fidelity digital twins, and real hardware. The main contributions are: 1. We provide a unified workflow for testing distributed algorithms on simple models, high-fidelity digital twins, and hardware from the same workstation, addressing multi-stage validation needs (Mokhtarian et al., 2024). 2. Unlike testbeds limited to a single robot morphology, our approach can be extended to support a variety of dynamical model. This enables easy transitions between heterogeneous systems. 3. The framework uses the Single Program, Multiple Data (SPMD) paradigm (Flynn, 1966; Foster, 1995; Gropp et al., 2014) to emulate decentralised execution, with each worker acting as an independent agent using local states and neighbour communication. This enables scalable algorithm-in-the-loop testing on a single multi-core workstation, avoiding complex networking while reducing simulation time through parallel execution. 4. We adapt and incorporate a Crazyflie model (Kim, 2017; Förster, 2015) along with hardware APIs, enabling consistent testing with simple model, high-fidelity model, and physical UAVs. With this approach only the actuation block changes while the rest of the distributed controller remains identical, fulfilling modularity. 5. We demonstrate the utility of the framework through a distributed position-swapping example and release the full code111https://tinyurl.com/smart-city-dcffor reproducibility and further development. Overall, we provide an end-to-end prototyping pipeline that significantly reduces testing time, lowers hardware requirements, and improves reproducibility. 1.3 Article Organization The remainder of this paper is organized as follows. Section 2 presents the proposed framework, including its workflow and computational setup for multi-fidelity testing of distributed optimization algorithms. Section 3 describes the implementation, software and hardware setup, and the four-UAV position-swapping problem. Section 4 reports the results for the simple model, digital twin, and hardware platform. Finally, Section 5 concludes the paper and discusses future platform extensions. 2 Prototyping Framework We envision testing of distributed optimization algorithms with practical multi-robot deployments, by organizing the prototyping workflow into a structured sequence of decisions and execution stages, as illustrated in Figure 1. Figure 1: Proposed workflow for testing distributed control on multi-robot systems. The workflow involves selecting the network topology (e.g., fully connected, star, mesh, ring etc.), execution mode (synchronous or asynchronous), and a controller or algorithm such as distributed pseudo-gradient or projected subgradient descent. It further supports three testing levels: simple dynamical models for rapid prototyping, high-fidelity digital twins, and physical hardware with associated APIs and localization modules. The modular architecture enables deployment across different models and platforms through a common computational interface. In Fig. 1, currently supported features are shown in green (with *), while planned extensions are shown in grey. Figure 1 also shows how these components integrate into a testing workflow at various levels of fidelity. First, any given algorithm should be tested on simple dynamical models (e.g., single or double integrators) to rapidly evaluate its performance and communication patterns. The second stage involves validating these algorithms with a digital twin (i.e. high-fidelity dynamic models) to capture underlying unmodeled dynamics (e.g., mechanics, aerodynamics), actuator limits, sensor noise, etc. Finally, the third stage transitions to real hardware, where the validated algorithms are deployed via the same algorithmic architecture, using an application layer to interface with multiple agents. All testing stages can be executed in a unified manner using a single multi-core workstation, with each core representing a logical agent in the multi-agent network, as shown in Figure 2. Framework employs the SPMD paradigm, where agents execute identical control programs using local states, decision variables, and neighbour information while exchanging only algorithm-required data. The approach reduces setup complexity and hardware cost while deploying parallel execution, logging, debugging, and integration with testing platforms. With this approach, we can emulate communication effects in code such as latency, packet loss, and bandwidth constraints for robustness evaluation. Figure 2: Distributed control setup using a single computer. 3 Methodology 3.1 Demonstration Problem For the framework demonstration, we consider a fleet of UAVs performing level flight while avoiding collisions and pursuing a shared objective. This yields an optimization problem coupled through the cost function (via the shared objective) and in the constraints (through collision avoidance), making it suitable for distributed computation. 3.1.1 System model: Consider a set I=1,…,NI=\1,…,N\, an indexed fleet of N UAVs. For the control design purposes, we assume that the UAVs share the same dynamics: p¨ix(t)=aix(t),p¨iy(t)=aiy(t) p^x_i(t)=a^x_i(t),\> p^y_i(t)=a^y_i(t), where pix(t),piy(t)p^x_i(t),p^y_i(t) denotes the ithi^th-UAV’s position in the (x,y)(x,y) plane respectively at time t. Similarly, we denote by vix(t),viy(t),aix(t),aiy(t)v^x_i(t),v^y_i(t),a^x_i(t),a^y_i(t) the ithi^th UAV velocity and acceleration components in the xyxy-plane. Following a zero order hold discretization with sample time tst_s, we obtain: zi[k+1]=[10ts0010ts00100001]zi[k]+[ts2/200ts2/2ts00ts]ai[k]z_i[k+1]= bmatrix1&0&t_s&0\\ 0&1&0&t_s\\ 0&0&1&0\\ 0&0&0&1 bmatrixz_i[k]+ bmatrixt_s^2/2&0\\ 0&t_s^2/2\\ t_s&0\\ 0&t_s bmatrixa_i[k] (1) where zi[k]=[pix[k]piy[k]vix[k]viy[k]]⊤∈ℝ4z_i[k]=[\>p^x_i[k]\>p^y_i[k]\>v^x_i[k]\>v^y_i[k]\>] ^4 denotes the state, ai[k]=[aix[k]aiy[k]]⊤∈ℝ2a_i[k]=[\>a^x_i[k]\>a^y_i[k]\>] ^2 is the control input and k is the discrete-time index. In the following discussion, AidA^d_i and BidB^d_i are used to denote the state and input matrices respectively in (1) for each agent. We assume each UAV starts at the initial position pi0=[pix0,piy0]⊤∈ℝ2p^0_i=[p^x_0_i,p^y_0_i] ^2 and has piT=[pixT,piyT]⊤∈ℝ2p^T_i=[p^x_T_i,p^y_T_i] ^2 as its target position . We re-derive (1) using the error vector ei[k]=zi[k]−[pixT,piyT, 0, 0]⊤e_i[k]=z_i[k]-[p^x_T_i,\>p^y_T_i,\>0,\>0\>] . The error dynamics for each agent can be written as: ei[k+1]=Aidei[k]+Bidai[k]e_i[k+1]=A^d_ie_i[k]+B^d_ia_i[k] (2) 3.1.2 Agent-wise optimization problem: Each agent i∈Ii∈ I solves a receding horizon control problem of horizon length H. Let ui=[ai[0]…ai[H−1]]iTu_i=[a_i[0]… a_i[H-1]]^T_i be the decision vector of agent i over the prediction horizon, where ui∈Ui⊆ℝ2Hu_i∈ U_i ^2H denotes the presence of local constraints UiU_i (to be defined in the sequel). Thus, each agent i∈Ii∈ I solves: minui∈Ui _u_i∈ U_i~ Ji(ui,σ(u)) J_i(u_i,σ(u)) subject to Aiui+∑j∈I∖iAjuj≤∑i∈Ibi A_iu_i+Σ _j∈ I \i\A_ju_j≤Σ _i∈ Ib_i (3) where j∈I∖ij∈ I \i\ represents all agents except agent i and Ji(ui,σ(u))J_i(u_i,σ(u)) is the finite horizon cost which each agent i∈Ii∈ I seeks to minimize. This cost is coupled via the aggregative variable σ(u):=1N∑i∈Iϕi(ui)σ(u):= 1N _i∈ I _i(u_i), which depends on all agents’ decisions uiu_i, formed by the aggregation rule ϕi(⋅) _i(·). Let po=[pox,poy]⊤p_o=[p_o^x,p_o^y] denote the common target position and do=‖1N∑z=1Npz[H]−po‖2d_o= \| 1N _z=1^Np_z[H]-p_o \|_2 the distance between the predicted swarm centroid at horizon step H and pop_o. We define the local cost function as follows: Ji(ui,σ(u))=∑l=0H−1ei[l]TQiei[l]+ui[l]TRiui[l] J_i(u_i,σ(u))= _l=0^H-1e_i[l]^TQ_ie_i[l]+u_i[l]^TR_iu_i[l] +βei[H]TPiei[H]+paNdo2 +β\,e_i[H]^TP_ie_i[H]+ p_aNd_o^2 (4) where the first two terms in (4) represents the running cost with weighting matrices Qi≻0Q_i 0 and Ri≻0R_i 0, while βPi≻0β P_i 0 for third term defines the terminal penalty. The final term, weighted by pa/Np_a/N, penalizes the swarm-centroid deviation relative to individual target positions. The local constraint set UiU_i captures the bounded UAV acceleration inputs: Ui=ui:−aimax≤ai[k]≤aimax,k=0,…,H−1 U_i=\u_i:-a_i ≤ a_i[k]≤ a_i ,\;k=0,…,H-1\ (5) The agent-wise optimization is also subject to coupled constraints introduced for collision avoidance. This is defined for each pair of agents via the discrete-time control barrier function. For i,j∈Ii,j∈ I with i≠ji≠ j, let δpi,j[k]=[pxi[k]pyi[k]]T−[pxj[k]pyj[k]]Tδ p_i,j[k]=[p_x_i[k]\>p_y_i[k]]^T-[p_x_j[k]\>p_y_j[k]]^T and a set S, being the superlevel set of a map hij:⊂ℝ2→ℝh^j_i:P ^2 S=δpi,j[k]∈⊂ℝn:hij(δpi,j[k])≥0.S=\δ p_i,j[k] ^n:h^j_i(δ p_i,j[k])≥ 0\. The set S denotes the safe set within which the agents’ trajectories must remain to avoid collisions. To ensure collision avoidance: (i)hij(δpi,j[0])≥0, (i)~h^j_i(δ p_i,j[0])≥ 0, (ii)∃ui[k],uj[k] such that ∀k∈ℕ∪0, (i)~∃\>u_i[k],u_j[k] such that ∀ k ∪\0\, (6) hij(δpi,j[k+1])−hij(δpi,j[k])≥−γcbf(hij(δpi,j[k])). 17.07182pth^j_i(δ p_i,j[k+1])-h^j_i(δ p_i,j[k])≥- _cbf(h^j_i(δ p_i,j[k])). The function hij:→ℝh^j_i:P is said to be a discrete-time exponential CBF and the set S is invariant, or safe, along the trajectories of (2) driven in closed loop by ui[k],uj[k]u_i[k],u_j[k] in (3.1.2), as defined in Agrawal and Sreenath (2017). The following candidate control barrier functions are used for (i,j)∈I,i≠j(i,j)∈ I,i≠ j: hij h_i^j (δpi,j[k])=|pxi[k]−pxj[k]|r1+|pyi[k]−pyj[k]|r2−1 (δ p_i,j[k])= p_x_i[k]-p_x_j[k] r_1+ p_y_i[k]-p_y_j[k] r_2-1 (7) =|eix[k]−ejx[k]+δpxi,jd|r1+|eiy[k]−ejy[k]+δpyi,jd|r2−1. = e_i^x[k]-e_j^x[k]+δ p^d_x_i,j r_1+ e_i^y[k]-e_j^y[k]+δ p^d_y_i,j r_2-1. The function hij(δpi,j[k])≥0h_i^j(δ p_i,j[k])≥ 0 guarantees no collision between UAVs i and j at time k. Parameters r1,r2>0r_1,r_2>0 are the norm-1 unsafe radii, delimiting the UAV’s body in the x and y directions. These conditions are then transformed into matrices AiA_i, AjA_j and vectors bib_i defined in (3). For the full derivation and characterization of AiA_i, AjA_j and bib_i, we refer the reader to do Nascimento et al. (2023). 3.1.3 Distributed optimization: Given our distributed safety requirements, the coupled constraints set the stage for a generalized games setting: typically, primal-dual iterative methods are then used, in which the Nash Equilibrium (NE) of the game is reached asymptotically. To enhance the real-time deployment potential, and effective peer-to-peer communication, we resort to a technique based on gradient tracking by Carnevale et al. (2024). This technique is able to uncouple the cost and constraints by introducing two local variables on each agent, being able to track asymptotically the aggregative cost and coupled constraints in each agent. Different from do Nascimento et al. (2023), in this work we distribute the computation among different computer cores and study its high fidelity deployment numerically and in the hardware. For the full algorithm definitions and derivations, we refer the reader to Carnevale et al. (2024). We highlight that the platform uses a modular distributed optimization scheme, and could be readily substituted by other algorithm of choice, such as the Alternating Direction Method of Multipliers (ADMM), a popular distributed optimization benchmark technique. 3.2 Software Setup The framework was implemented using the MATLAB® Parallel Computing Toolbox, where agents execute in parallel within an SPMD block and each computational engine corresponds to a worker. Algorithm 1 summarizes the software workflow. For UAV prototyping, the double-integrator dynamics in (1) were used for initial testing. The distributed position-swapping routine from do Nascimento et al. (2023); Carnevale et al. (2024) was executed within the SPMD framework, where each worker represented a UAV agent maintaining local states, solving the distributed optimization problem, and exchanging neighbour information. Algorithm-1 Pseudocode for distributed control using SPMD ⬇ % initialize variables e.g. numUAVs,maximum number of iteration steps, etc. setup parameters spmd (numUAVs) % Each worker represents a UAV agent id = spmdIndex; % spmdIndex is the agent’s number. % 0. initialize state and define neighbours for k = 1:steps % iterations % 1. Communicate Exchange information with neighbours using spmdSendReceive % 2. Compute local control action using chosen iterative algorithm % 3. Sense and Act Measure state and apply action to the system i.e. simple model/ high fidelity model/ real hardware % 4. Synchronization between agents (Optional step- depends on type of routine % 5. Log data or send to base station end end % post processing and visualizations For high-fidelity testing, the distributed routine was integrated with a Simulink model adapted from Kim (2017), based on the non-linear Crazyflie 2.0 dynamics identified in Förster (2015). The model includes modified low-level PID controllers. Additional interfaces were added to integrate with the SPMD framework. Using the MATLAB 2024a “Simulation object” feature222https://tinyurl.com/simulink-simulation, each worker executes its local UAV simulation while exchanging distributed control information with neighbouring workers. This modular structure preserves the communication, synchronization, and logging routines while routing agent dynamics through either simplified, high-fidelity, or hardware-integrated models. The complete distributed control framework codebase is publicly available at Github333https://tinyurl.com/smart-city-dcf, including utilities for high-fidelity simulation, hardware experiments, data collection, and visualization. 3.3 Hardware Setup and Integration For UAV experiments, Crazyflie 2.1 drones were used which comes with its Python API444https://github.com/bitcraze/crazyflie-lib-python for wireless communication and telemetry acquisition. The hardware setup is shown in Fig. 3. Localization is provided by the Lighthouse-V2 infrared motion-tracking system, where onboard photodiodes detect laser sweeps from multiple base stations to triangulate the UAV pose with approximately 1 cm accuracy (Taffanel et al., 2021). This proved sufficient for our distributed control experiments. Communication with the UAVs is established using Crazyradio555https://w.bitcraze.io/products/crazyradio-pa/ USB dongles operating in the 2.4 GHz ISM band. Although a single dongle can support multiple UAVs, one Crazyradio was assigned per agent to reduce latency and communication contention. The interface is managed through the Crazyflie Python API and integrated into MATLAB through Python calls. The radios support data rates up to 2 Mbps and the Lighthouse-V2 system operates at approximately 30 Hz, which were sufficient for waypoint updates to the onboard controllers. The full command chain works as follows: after each distributed optimization step (Algorithm 1, Step 2), every worker transmits planar velocity references to its assigned Crazyflie. Using Lighthouse-V2 and onboard sensor measurements, the UAV estimates its state via an extended Kalman filter. The onboard PID velocity and attitude controllers then convert these references into motor PWM commands to track references. Updated state estimates are transmitted back to workstation through the Crazyradio interface for the next distributed control iteration. Figure 3: Hardware setup for distributed control of multi-UAV systems using a single multi-core computer. Figure 4: Trajectories for (left) double integrator, (center) high-fidelity, and (right) UAV experiments. 4 Results and Discussion We evaluate a four-UAV position-swapping task using a synchronous distributed controller over a fully connected communication network. Trajectories are compared across the double-integrator model, the Crazyflie high-fidelity model, and Crazyflie 2.1 hardware using the Lighthouse-V2 localization system. The initial positions are p10=[0,1]Tp^0_1=[0,1]^T, p20=[0,−1]Tp^0_2=[0,-1]^T, p30=[1,0]Tp^0_3=[1,0]^T, and p40=[−1,0]Tp^0_4=[-1,0]^T, with targets defined as p1T=p20p^T_1=p^0_2, p2T=p10p^T_2=p^0_1, p3T=p40p^T_3=p^0_4, and p4T=p30p^T_4=p^0_3. The common tracking point is po=[0,0]Tp_o=[0,0]^T with fleet weight pa=1p_a=1. The sample time is ts=0.2st_s=0.2\,s and the acceleration bounds are amax=2m/s2a =2\,m/s^2 in both planar directions. The optimization parameters are H=3H=3, Qi=diag(5,5,5,5)Q_i=diag(5,5,5,5), Ri=diag(2,2,2,2)R_i=diag(2,2,2,2), and PiP_i equal to the Riccati solution of the unconstrained LQR problem without the aggregative term. Additionally, β=1.5β=1.5, γcbf=0.1 _cbf=0.1, and safety radii r1=0.5mr_1=0.5\,m and r2=0.25mr_2=0.25\,m. Figure 4 compares the controller across all platform levels. Circles mark the trajectory endpoints, with the opposite ends indicating the UAV starting positions. The similar trajectories suggest that the double-integrator model is a reasonable motion-planning approximation, as low-level PIDs quickly correct tracking errors. The high-fidelity model, however, captures overshoot also observed in hardware, making it useful for parameter tuning. Early termination, needed for real-time deployment since the algorithm reaches the NE only asymptotically, weakens safety guarantees. The stopping criterion requires the maximum of the primal and dual differences between consecutive iterates, e.g., p and p+1, to fall below a heuristic threshold. Despite this and the model mismatch, actual collisions were rarely observed. The high-fidelity model shows two agents briefly entering each other’s safety region and overshoot when reaching the target point, later confirmed in hardware but absent in the point-mass model. No collisions occurred, partly due to the generous CBF safety radii. A recorded demonstration is available online666https://youtu.be/xZksNosFYro?si=wq258DJx4m99w_6R. Figure 5 summarizes the computational performance of the distributed MPC controller across all framework levels. The boxplots show computation times measured around Step 2 of Algorithm 1 for the point-mass model, high-fidelity Crazyflie simulator, and Crazyflie 2.1 hardware. After removing one extreme outlier using a 10×IQR10×IQR criterion, all platforms exhibit similar distributions with medians well below the 200ms200\,ms real-time budget. The hardware results closely match the point-mass model in both median and variability, supporting the use of simplified models for early-stage algorithm testing. The high-fidelity model is consistently slower since each worker simulates its local UAV dynamics. Overall, the results demonstrate real-time feasibility of the distributed MPC scheme across all testing levels. 5 Conclusion This paper presented a prototyping framework for distributed control of multi-robot systems, combining parallel computing with a modular structure that supports an expanding library of simulation models and hardware APIs. Integration with Crazyflie quadrotor hardware demonstrated efficient single-workstation validation of distributed algorithms without complex networking infrastructure or dedicated multi-robot testbeds. Figure 5: Computation time distribution for four agents. The approach has limitations: wireless interference, radio contention, and physical timing jitter cannot be fully reproduced in single-host emulation and must be assessed on real hardware. Nevertheless, the framework provides a fast, low-cost, and flexible environment for iterative development, reducing the effort required for performance testing. Future work will expand the model and hardware libraries, enabling broader comparisons across fidelity levels and robotic platforms. The framework may also be used to study communication effects and adversarial disturbances, supporting robustness analysis in multi-robot distributed control. References A. Agrawal and K. Sreenath (2017) Discrete Control Barrier Functions for Safety-Critical Control of Discrete Systems with Application to Bipedal Robot Navigation. In Robotics: Science and Systems XIII, (en). External Links: ISBN 978-0-9923747-3-0, Document Cited by: §3.1.2. J. Blumenkamp, A. Shankar, M. Bettini, J. Bird, and A. Prorok (2024) The Cambridge RoboMaster: An Agile Multi-Robot Research Platform. arXiv. Note: arXiv:2405.02198 [cs] External Links: Document Cited by: §1.1. G. Carnevale, F. Fabiani, F. Fele, K. Margellos, and G. Notarstefano (2024) Tracking-Based Distributed Equilibrium Seeking for Aggregative Games. IEEE Transactions on Automatic Control 69 (9), p. 6026–6041. Note: Conference Name: IEEE Transactions on Automatic Control External Links: ISSN 1558-2523, Document Cited by: §3.1.3, §3.2. H. Deng, Q. Fu, Q. Quan, K. Yang, and K. Cai (2020) Indoor Multi-Camera-Based Testbed for 3-D Tracking and Control of UAVs. IEEE Transactions on Instrumentation and Measurement 69 (6), p. 3139–3156. External Links: ISSN 1557-9662, Document Cited by: §1.1. A. A. do Nascimento, A. Papachristodoulou, and K. Margellos (2023) A Game Theoretic Approach for Safe and Distributed Control of Unmanned Aerial Vehicles. In 2023 62nd IEEE Conference on Decision and Control (CDC), p. 1070–1075. Note: ISSN: 2576-2370 External Links: Document Cited by: §3.1.2, §3.1.3, §3.2. M.J. Flynn (1966) Very high-speed computing systems. Proceedings of the IEEE 54 (12), p. 1901–1909. External Links: ISSN 1558-2256, Document Cited by: item 3. J. Förster (2015) System Identification of the Crazyflie 2.0 Nano Quadrocopter. Bachelor Thesis, ETH Zurich, (en). External Links: Document Cited by: item 4, §3.2. I. Foster (1995) Designing and Building Parallel Programs. Addison-Wesley Longman Publishing Co., Inc. (en). Cited by: item 3. W. Gropp, E. Lusk, and A. Skjellum (2014) Using MPI: Portable Parallel Programming with the Message-Passing Interface. MIT Press, Cambridge, UNITED STATES. External Links: ISBN 978-0-262-32660-5 Cited by: item 3. A. Jamshidpey, M. Wahby, M. Allwright, W. Zhu, M. Dorigo, and M. K. Heinrich (2025) Centralization vs. decentralization in multi-robot sweep coverage with ground robots and uavs. Artificial Life and Robotics, p. 1–22. Cited by: §1. W. Kim (2017) Crazyflie quadcopter simulation using simmechanics. Note: MATLAB Central File Exchange External Links: Link Cited by: item 4, §3.2. E. T. (. T. King (2004) Distributed coordination and control experiments on a multi-UAV testbed. Thesis, Massachusetts Institute of Technology, (en_US). Note: Accepted: 2005-09-27T18:55:19Z Cited by: §1.1. A. Mokhtarian, J. Xu, P. Scheffe, M. Kloock, S. Schäfer, H. Bang, V. Le, S. Ulhas, J. Betz, S. Wilson, et al. (2024) A survey on small-scale testbeds for connected and automated vehicles and robot swarms: a guide for creating a new testbed. IEEE Robotics & Automation Magazine. Cited by: item 1, §1. G. Notarstefano, I. Notarnicola, A. Camisa, et al. (2019) Distributed optimization for smart cyber-physical networks. Foundations and Trends® in Systems and Control 7 (3), p. 253–383. Cited by: §1.1. L. Paull, J. Tani, H. Ahn, J. Alonso-Mora, L. Carlone, M. Cap, Y. F. Chen, C. Choi, J. Dusek, Y. Fang, et al. (2017) Duckietown: an open, inexpensive and flexible platform for autonomy education and research. In 2017 IEEE International Conference on Robotics and Automation (ICRA), p. 1497–1504. Cited by: §1.1. D. Pickem, P. Glotfelter, L. Wang, M. Mote, A. Ames, E. Feron, and M. Egerstedt (2017) The Robotarium: A remotely accessible swarm robotics research testbed. In 2017 IEEE International Conference on Robotics and Automation (ICRA), p. 1699–1706. External Links: Document Cited by: §1.1. M. Rubenstein, C. Ahler, N. Hoff, A. Cabrera, and R. Nagpal (2014) Kilobot: a low cost robot with scalable operations designed for collective behaviors. Robotics and Autonomous Systems 62 (7), p. 966–975. Cited by: §1.1. R. Shakeri, M. A. Al-Garadi, A. Badawy, A. Mohamed, T. Khattab, A. K. Al-Ali, K. A. Harras, and M. Guizani (2019) Design challenges of multi-uav systems in cyber-physical applications: a comprehensive survey and future directions. IEEE Communications Surveys & Tutorials 21 (4), p. 3340–3385. Cited by: §1. O. Shorinwa, T. Halsted, J. Yu, and M. Schwager (2024) Distributed optimization methods for multi-robot systems: part 2—a survey. IEEE Robotics & Automation Magazine 31 (3), p. 154–169. Cited by: §1.1. A. Stager, L. Bhan, A. Malikopoulos, and L. Zhao (2018) A Scaled Smart City for Experimental Validation of Connected and Automated Vehicles. IFAC-PapersOnLine 51 (9), p. 130–135 (en). External Links: ISSN 24058963, Document Cited by: §1.1. G. Stomberg, H. Ebel, T. Faulwasser, and P. Eberhard (2023) Cooperative distributed mpc via decentralized real-time optimization: implementation results for robot formations. Control Engineering Practice 138, p. 105579. Cited by: §1.1. G. Stomberg, A. Engelmann, M. Diehl, and T. Faulwasser (2025) Decentralized real-time iterations for distributed nmpc. IEEE Trans. on Automatic Control. Cited by: §1.1. A. Taffanel, B. Rousselot, J. Danielsson, K. McGuire, K. Richardsson, M. Eliasson, T. Antonsson, and W. Hönig (2021) Lighthouse Positioning System: Dataset, Accuracy, and Precision for UAV Research. arXiv. Note: arXiv:2104.11523 [cs] External Links: Document Cited by: §3.3. S. Wilson, P. Glotfelter, L. Wang, S. Mayya, G. Notomista, M. Mote, and M. Egerstedt (2020) The Robotarium: Globally Impactful Opportunities, Challenges, and Lessons Learned in Remote-Access, Distributed Control of Multirobot Systems. IEEE Control Systems Magazine 40 (1), p. 26–44. External Links: ISSN 1941-000X, Document Cited by: §1.1. T. Yang, X. Yi, J. Wu, Y. Yuan, D. Wu, Z. Meng, Y. Hong, H. Wang, Z. Lin, and K. H. Johansson (2019) A survey of distributed optimization. Annual Reviews in Control 47, p. 278–305. Cited by: §1.1.