PaperlayerView exact revision on arXiv

Source-linked AI summary

Evolution of Collective Behaviors for a Real Swarm of Aquatic Surface Robots

Miguel Duarte, Vasco Costa, Jorge Gomes, Tiago Rodrigues, Fernando Silva, Sancho Moura Oliveira, Anders Lyhne Christensen

arXiv:1511.03154v2cs.RO

TL;DR

Swarm robotics has promise for collective tasks, but most studies had remained in simulation or controlled laboratories, and synthesizing swarm behavior is challenging. This paper evolves distributed controllers in simulation for aquatic surface robots and evaluates them in an uncontrolled real environment. The authors report successful operation across canonical tasks and a combined environmental-monitoring task, with real-swarm behavior and performance similar to simulation.

  • Problem

    Most swarm robotics studies were conducted in simulation or controlled laboratories, while deriving local robot rules for a desired global behavior remains challenging.

    Source: §1 · pp. 1–2

  • Method

    The authors evolve distributed controllers for aquatic robots in simulation and combine evolved behaviors for an environmental-monitoring task.

    Source: §1 · p. 2

  • Results

    The evolved controllers operated in an uncontrolled real environment across four canonical tasks, with behavior and performance similar to simulation and a combined monitoring task demonstrated.

    Source: §1 · p. 2, §9 · p. 22

  • Takeaways & Limitations

    The work demonstrates the potential of swarm and evolutionary robotics for real-world multirobot scenarios, including maritime tasks requiring distributed sensing.

    Source: §9 · p. 22

    As task complexity increases, designing effective fitness functions becomes difficult, and standard evolutionary techniques can face bootstrapping and deception issues.

    Source: §8.1 · p. 18

Abstract

from arXiv · show

Swarm robotics is a promising approach for the coordination of large numbers of robots. While previous studies have shown that evolutionary robotics techniques can be applied to obtain robust and efficient self-organized behaviors for robot swarms, most studies have been conducted in simulation, and the few that have been conducted on real robots have been confined to laboratory environments. In this paper, we demonstrate for the first time a swarm robotics system with evolved control successfully operating in a real and uncontrolled environment. We evolve neural network-based controllers in simulation for canonical swarm robotics tasks, namely homing, dispersion, clustering, and monitoring. We then assess the performance of the controllers on a real swarm of up to ten aquatic surface robots. Our results show that the evolved controllers transfer successfully to real robots and achieve a performance similar to the performance obtained in simulation. We validate that the evolved controllers display key properties of swarm intelligence-based control, namely scalability, flexibility, and robustness on the real swarm. We conclude with a proof-of-concept experiment in which the swarm performs a complete environmental monitoring task by combining multiple evolved controllers.

1 BioMachines Lab, Lisbon, Portugal

The paper addresses the difficulty of designing decentralized swarm behavior and the lack of demonstrations in uncontrolled real environments, presenting evolved control on aquatic robots as a step toward such applications. It situates this work against prior simulation- and laboratory-based studies and the reality gap in transferring evolved controllers.

1 Introduction

  • Swarm systems coordinate through decentralized local interactions and may offer robustness, flexibility, and scalability, but real-robot demonstrations had remained largely confined to controlled laboratories.The introduction also motivates maritime swarms for tasks requiring distributed sensing and fault tolerance.Source: §1 · pp. 1–2
  • For the first time, evolved distributed control is demonstrated in an uncontrolled real-world environment on up to ten aquatic robots performing four collective tasks and a combined monitoring task.The paper also reports testing scalability and robustness of the controllers in a real environment.Source: §1 · p. 2

2.1 Swarm Robotics

  • Real-world swarm robotics remains difficult because designing individual rules for complex tasks and uncontrolled environments is challenging.The paper identifies behavioral-control design as a major obstacle to moving beyond abstract tasks and controlled settings.Source: §2.1 · p. 3

2.2 Evolutionary Synthesis of Control for SRS

  • Evolutionary robotics can synthesize self-organized control from a high-level task metric, but most studies were simulated and transferring controllers risks a reality gap.Prior work proposed approaches including conservative noise, minimal simulation, transferability evaluation, and reactive behavior.Source: §2.2 · pp. 3–4

2.3 Aquatic Robots

  • Aquatic multi-robot systems had few realized examples, while prior evolved aquatic control studies covered a single simulated robot or centralized control for four simulated robots.A cited distributed-monitoring system used a central computer, limiting remote deployment and potential scaling.Source: §2.3 · pp. 4–5

3 Methodology

The methodology uses a simulation platform to evolve and assess controllers before transferring them to real robots. It reduces simulation-to-hardware implementation differences and models sensor and actuator uncertainty using observations from the real robots.

3 Methodology

  • The methodology synthesizes controllers for four tasks and assesses steps intended to support successful transfer to the real swarm.The process is summarized in Figure 2.Source: §3 · p. 5, Figure 2 · p. 5

1. Definition of the simulation platform

  • The two-dimensional simulator represents robots as circles with position and heading, models motion from water measurements, and omits fluid dynamics to keep evolutionary evaluation computationally feasible.A shared middle layer lets the same code base and controllers run in simulation and on real robots, minimizing implementation differences.Source: §1 · p. 5

2. Configuration of the robots’ sensors and actuators

  • Robot sensors provide distances to known objects in equal-sized segments around each robot, derived from GPS locations, heading, and position.Simulation adds observation-based noise for parameter variation, per-step sensor and actuator errors, and water currents to reflect real-robot stochasticity.Source: §2 · pp. 5–6

3. Definition of the fitness evaluation

Task-specific fitness functions evolved neural-network controllers in simulation, with variation in trial conditions to reduce overfitting. Homing, dispersion, and monitoring solutions emerged within 100 generations, while clustering required up to 400.

3. Definition of the fitness evaluation

  • Controllers were scored across 10 simulation trials with varied robot numbers, starting poses, and spatial-entity locations, then evolved over 10 runs and post-evaluated in 100 simulations per generation.NEAT optimized neural-network connection weights and topology.Source: §3 · p. 6

5. Selection of evolved controllers

  • For each task, the three highest-scoring best-of-run controllers were tested in three real-robot experiments and compared with simulations using the same initial positions.The three candidates were selected from the ten best-of-run controllers.Source: §5 · p. 6

4.1 Robotic Platform

  • Experiments used up to ten GPS- and compass-equipped aquatic robots with Wi-Fi position broadcasts to test homing, dispersion, clustering, and monitoring.Each robot broadcast its position to neighboring robots every second.Source: §4.1 · p. 7, §4.2 · p. 8

Homing

  • Homing rewards reducing swarm-average waypoint distance, while fitness penalizes robot pairs closer than 3 m.The task requires waypoint navigation while avoiding collisions.Source: Homing · pp. 8–9

Dispersion

  • Dispersion controllers minimize the difference between nearest-neighbor distance and a 20 m target.The target distance is half the robots’ communication range.Source: Dispersion · p. 9

Clustering

  • Clustering controllers are scored by the number of clusters over time, with robots less than 7 m apart counted in the same cluster.Robots start randomly spread across a 100 m square and seek to form one cluster, though limited communication and an unbounded environment can leave multiple clusters.Source: Clustering · p. 9

Area Monitoring

  • Monitoring controllers maximize coverage over time on a 1 m² grid across varied geo-fences, with 5 m robot coverage and 100 s linear cell-value decay.Source: Area Monitoring · p. 10

4.3 Evolutionary Results

  • Homing, dispersion, and monitoring solutions emerged within 100 generations, while clustering required up to 400 generations.The authors attribute clustering’s longer evolution to the challenge of forming one cluster from random starting positions in an unbounded environment.Source: §4.3 · p. 10

5 Transferring Evolved Controllers to Real Robots

The evolved controllers transferred to real aquatic robots across homing, dispersion, clustering, and monitoring, but transfer quality varied by task. Monitoring outperformed simulation, while dispersion exposed a motion-model mismatch.

5 Transferring Evolved Controllers to Real Robots

  • The top three controllers for each task were tested in three independent experiments per controller at Parque das Nações, a 300 m × 190 m waterbody with wind speeds of 15–40 km/h.Average robot drift speed was 0.2 m/s, twice the value used in simulation.Source: §5 · pp. 10–11

5.1 Homing

  • All three homing controllers reached successive waypoints while avoiding collisions, with real and simulated robots arriving at similar times and maintaining an average distance of around 10 m.Waypoints changed every 60 seconds during the four-minute experiment.Source: §5.1 · pp. 11–12, Figure 7 · p. 12

5.2 Dispersion

  • Real-robot dispersion exposed a reality gap: Controller 3 had 2 m average error, while Controllers 1 and 2 performed considerably worse despite similar simulation performance.The real robots’ low-speed settling circles were larger than those in simulation, moving them farther from their positions.Source: §5.2 · p. 13, Figure 8 · p. 13

5.3 Clustering

  • Clustering controllers generally crossed the reality gap, but outcomes varied with initial positions; Controller 1 formed one cluster in two of three samples.Controller 2 performed worse than in simulation, while Controller 3 performed similarly.Source: §5.3 · pp. 13–14, Figure 9

5.4 Area Monitoring

  • Monitoring succeeded in all three areas and performed significantly better on real robots than in simulation, although L-shaped coverage was uneven.Robots that drifted beyond the boundaries because of wind, currents, GPS inaccuracies, or nearby robots quickly returned and resumed monitoring.Source: §5.4 · pp. 14–15, Figure 10 · p. 15

6 Scalability and Robustness

Real-swarm tests found that dispersion and clustering generally worked across swarm sizes, while controllers adapted to added robots and changes in monitoring-swarm composition. Clustering outcomes were constrained by limited communication range.

6.1 Scalability

  • Dispersion worked across swarm sizes, with a slight performance decrease at six robots; clustering formed one cluster in all three four-robot samples and two of three samples at six and eight robots.Limited communication range can leave subsets disconnected and produce multiple clusters.Source: §6.1 · p. 16, Figure 11 · p. 16

6.2 Robustness

  • Dispersion controllers adjusted when four new robots joined: the original group moved farther apart, and distance error decreased as all eight robots continued dispersing.The added robots first moved to the swarm center, then began dispersing after 180 seconds.Source: §6.2 · p. 16, Figure 12 · p. 17
  • Monitoring performance decreased after four of eight robots were removed at five minutes and increased after two were added at ten minutes, indicating approximately linear scaling with robot count and robustness to composition changes.The test ran for 15 minutes on a 100 × 100 m square area.Source: §6.2 · pp. 16–17

7 Multi-controller Mission

The swarm combined evolved homing, dispersion, monitoring, and clustering behaviors in a sequential water-temperature sampling mission. As robots covered the area, the temperature estimate became more accurate.

7 Multi-controller Mission

  • The mission combined five sequential subtasks: travel from base to the area center, disperse, monitor, aggregate, and return to base.The robots used homing for both travel legs and clustering to aggregate.Source: §7 · pp. 17–18
  • Robots began collecting temperature data after reaching the monitoring area’s center, before the dispersion subtask.Each behavior ran for a predefined amount of time.Source: §7 · p. 18
  • By the end of monitoring at 360 seconds, estimated standard deviation of temperature-estimation error was below 0.14 across the area except in corners, where it was below 0.24.The estimate’s uncertainty decreased as the robots covered the area.Source: §7 · p. 18

8 Discussion

The experiments support evolutionary computation as a viable way to synthesize swarm control for uncontrolled environments. The discussion examines evolutionary design and transfer challenges, reports robustness, flexibility, and scalability, and identifies remaining barriers to practical use.

8.1 Evolutionary Process

  • Fitness functions became difficult to design for complex sequential tasks such as temperature monitoring, where standard evolutionary techniques can face bootstrapping and deception issues.The multi-controller mission instead combined behaviors that each ran for a predefined time.Source: §8.1 · p. 18

8.2 Transferability of Controllers

  • Controllers generally crossed the reality gap, but speed differences of up to 20% and larger-than-simulated dispersion circles reduced performance in some cases.The authors attribute the motion mismatch to the simplified simulation physics and propose improving dynamics or selecting for transferable motion patterns.Source: §8.2 · p. 20

8.3 Swarm Behaviors

  • The swarm showed robustness when robots were removed during monitoring, and sporadic hardware failures did not impact group behavior across tasks.The authors also observed flexibility through different coordination strategies and sequential combination of evolved controllers, and scalability across four to eight robots and proportional monitoring coverage.Source: Robustness · p. 20

8.4 Robustness to Hardware Faults

  • Hardware faults not modeled during evolution—including motors temporarily stopping, reduced motor power, and erroneous GPS readings—did not compromise overall swarm performance.The slowest-to-fastest robot speed difference reached 0.6 m/s in the real monitoring swarm, versus 0.1 m/s in simulation.Source: §8.4 · pp. 20–21

8.5 Towards Real World Applications

  • Practical applications may require integrating mission-specific sensor data, which must be made meaningful to neural controllers and accurately modeled in simulation.The experiments relied mostly on robot and neighbor positions for behavioral decisions.Source: §8.5 · p. 21
  • Real-world use may require fault detection, fault tolerance, or online adaptation to ensure faulty individuals do not compromise the swarm.The experiments showed some tolerance to individual faults, but the authors note that applications may require stronger guarantees.Source: §8.5 · p. 21
  • Robots need to handle external factors such as nearby vessels while ensuring safety and regulatory compliance.Combining fail-safe behaviors with mission controllers may require evolved and manually programmed behaviors.Source: §8.5 · p. 21
  • Rapidly producing or modifying behaviors remains difficult; the study arranged evolved behaviors sequentially for monitoring, while hierarchical composition is proposed for future work.The authors identify controller composition as a route toward greater flexibility and modularity.Source: §8.5 · p. 21

9 Conclusion

Evolved controllers performed similarly in simulation and on real aquatic robots in an uncontrolled environment, displaying robustness, flexibility, and scalability. The swarm also combined simple behaviors to complete an environmental monitoring task.

9 Conclusion

  • Controllers for homing, clustering, dispersion, and monitoring showed similar behaviors and performance on real robots as in simulation, with robustness, flexibility, and scalability.The experiments used a swarm of small aquatic surface robots in an uncontrolled environment.Source: §9 · p. 22
  • The swarm combined simple evolved behaviors to carry out a complete environmental monitoring task.Source: §9 · p. 22
  • The study contrasts with prior work largely confined to simulation or controlled laboratories by demonstrating evolved swarm control in a realistic, uncontrolled environment.The authors conclude that swarm robotics and evolutionary robotics may be viable approaches for real-world multirobot control.Source: §9 · p. 22

Supporting Information

The supporting information includes an experimental-details document and two videos covering adaptive monitoring and a multi-controller mission.

Supporting Information

Experimental Details

The platform comprises ten small, inexpensive aquatic robots with onboard control and wireless communication. Simulation parameters draw on measurements of the real robots and model noise and delay conditions.

Robotic Platform

  • The platform comprised ten 60 cm robots costing approximately 300 EUR each, built as differential-drive monohull boats with Raspberry Pi 2 control units.The robots used digital manufacturing and off-the-shelf hardware to keep costs low.Source: Robotic Platform · p. 28
  • Robots broadcast identification, position, and orientation every second over an ad-hoc Wi-Fi network, with empirical communication range up to 40 m.Source: Robotic Platform · p. 28

Experimental Parameters

Loading 1511.03154v2…