Collision avoidance using a model of the locust LGMD neuron

https://doi.org/10.1016/S0921-8890(99)00063-9Get rights and content

Abstract

The lobula giant movement detector (LGMD) system in the locust responds selectively to objects approaching the animal on a collision course. In earlier work we have presented a neural network model based on the LGMD system which shared this preference for approaching objects.

We have extended this model in order to evaluate its responses in a real-world environment using a miniature mobile robot. This extended model shows reliable obstacle detection over an eight-fold range of speeds, and raises interesting questions about basic properties of the biological system.

Introduction

For animals the ability to detect approaching objects is important for survival, serving both to prevent collisions as the animal moves and also to avoid capture by predators [30]. While the fate of a mobile robot is unlikely to involve being eaten the ability to avoid collisions is equally important. Traditionally robotic technology has involved active sensors, such as ultrasound and infra-red devices, or high-precision sensors, such as laser scanners, for the detection of obstacles [7]. In biology, however, many examples are found of systems which rely on visual information to accomplish this task, and these have been successfully applied to real-world robotic control tasks [8], [49]. Neurons tuned to respond to approaching objects have been identified in species as diverse as pigeons [42] and locusts [31], [38].

The lobula giant movement detector (LGMD), a large visual interneuron in the optic lobe of the locust [24], is one such neuron. Originally the LGMD was thought to be tuned to detect novel movement of small objects [35]. However, more recent work has shown that the LGMD responds most strongly to approaching objects [31], [38] and that it is tightly tuned to objects approaching the animal on a direct collision course [14]. Receding objects produce little or no response [31]. Approaching objects are distinguished by the LGMD using the increasing speed of edge movement and increasing length of the edges [40]. The response dynamics during the approach of an object are the subject of two conflicting reports, one showing that the spike rate increases continuously during approach [31], the other showing that the spike rate may peak before collision [9], [12]. At the time of writing, this conflict is unresolved.

The LGMD is believed to play a role in triggering escape jumps and steering responses during flight. This belief is supported by the strong connection of the LGMD to the descending contralateral movement detector (DCMD) neuron [23] which in turn makes connections with interneurons and motoneurons in the thoracic ganglia [5], [39]. Responses in the DCMD match responses in the LGMD one-for-one [23], [27].

A neural network model of the input circuitry of the LGMD was developed by Rind and Bramwell [29]. Based closely on the anatomy and physiology of the optic lobe, the network comprised three principal layers:

  • the input photoreceptive layer, which responded to changes in the image in order to detect the edges of moving objects;

  • a processing layer, where excitation passed retinotopically through the network while delayed inhibition spread laterally;

  • the output layer, which represented the LGMD, where the excitation and inhibition were combined.

In addition, a feed-forward pathway inhibited the output of the model LGMD during large changes in the image, such as those caused by ego-motion. This model displayed the same preference for approaching objects as the LGMD and revealed that, at least for simple stimuli, this preference results from a critical race between the excitation produced by the movement of an object’s edges and the delayed lateral flow of inhibition within the network. This model was subsequently extended to allow the network to respond to textured stimuli [3].

In this paper we evaluate the behaviour of the LGMD network in a real-world environment. We show that robust obstacle detection can be achieved using an insect-based solution which relies on vision. Our results have interesting implications for the understanding of the biological LGMD system and demonstrate the potential of our LGMD model for real-world obstacle detection applications.

Section snippets

Experimental apparatus

Experiments were conducted in the environment shown in Fig. 1. This comprised of small stacks of Duplo blocks of various colours (red, green or blue) within a white outer wall. At the widest point the blocks were separated by 40 cm and at the narrowest point by 33 cm (Fig. 1(b)). The spaces between blocks were approximately 5 cm. The floor of the environment was clear Perspex over a gray sheet of paper.

A Khepera mobile robot (K-Team A.G., Lausanne, Switzerland) was fitted with a monochrome pinhole

Results

The directional selectivity of both the LGMD neuron and our neural network model is illustrated in Fig. 9. During the approach of an object the spike rate from the LGMD increases whereas there is only a brief response to receding objects. This response is found over a wide range of speeds of movement [28] and is tightly coupled to objects moving on a direct collision course [14].

In our present study we ran two series of experiments. In the first set of experiments, using only reactive infra-red

Discussion

In this project we evaluated a biologically based model of the locust LGMD in behavioural terms using a mobile robot. We found that the model responded reliably as the robot approached obstacles in its path over an eight-fold range of speeds. This illustrates that robust obstacle avoidance can be achieved using an insect based solution which relies on vision. Our experiments provided new insights into the dynamic responses of our model, which will be investigated in future experiments using

Acknowledgements

The authors wish to thank the anonymous referees for their constructive comments.

Mark Blanchard is a postdoctoral researcher at the Institute of Neuroinformatics, University-ETH Zurich, Switzerland. He received both his first degree (Microelectronics and Software Engineering, 1993) and his doctorate (Neurobiology, 1998) at the University of Newcastle upon Tyne, UK. His current research uses mobile robots to study aspects of insect vision which cannot be studied by traditional physiological and simulation methods.

References (50)

  • N. Franceschini, J.M. Pichon, C. Blanes, Bionics of visuo-motor control, in: T. Gomi (Ed.), Evolutionary Robotics: From...
  • F. Gabbiani et al.

    Computation of object approach by a wide-field motion-sensitive neuron

    Journal of Neuroscience

    (1999)
  • J.J. Gibson, The Ecological Approach to Visual Perception, Houghton Mifflin, Boston, MA,...
  • R.R. Harrison, C. Koch, An analog VLSI model of the fly elementary motion detector, in: M.I. Jordan, M.J. Kearns, S.A....
  • N. Hatsopoulos et al.

    Elementary computation of object approach by a wide-field visual neuron

    Science

    (1995)
  • G. Indiveri, P. Verschure, Autonomous vehicle guidance using analog VLSI neuromorphic sensors, in: W. Gerstner, A....
  • S.J. Judge et al.

    The locust DCMD, a movement-detecting neurone tightly tuned to collision trajectories

    Journal of Experimental Biology

    (1997)
  • J. Kramer, G. Indiveri, Neuromorphic vision sensors and preprocessors in system applications, in: Proceedings of the...
  • H.G. Krapp, F. Gabbiani, C. Koch, G. Laurent, Neuronal multiplication in the locust visual system, in: N. Elsner, R....
  • D. Lambrinos et al.

    An autonomous agent navigating with a polarized light compass

    Adaptive Behaviour

    (1997)
  • S. Liu, Silicon model of motion adaptation in the fly visual system, in: Proceedings of the Third UCSD–Caltech Joint...
  • M.A. Mahowald et al.

    The silicon retina

    Scientific American

    (1991)
  • R. Möller, D. Lambrinos, R. Pfeifer, T. Labhart, R. Wehner, Modeling ant navigation with an autonomous agent, in: R....
  • D.C. O’Carroll et al.

    Insect motion detectors matched to visual ecology

    Nature

    (1996)
  • M. O’Shea et al.

    The anatomy of a locust visual interneurone: The descending contralateral movement detector

    Journal of Experimental Biology

    (1974)
  • Cited by (0)

    Mark Blanchard is a postdoctoral researcher at the Institute of Neuroinformatics, University-ETH Zurich, Switzerland. He received both his first degree (Microelectronics and Software Engineering, 1993) and his doctorate (Neurobiology, 1998) at the University of Newcastle upon Tyne, UK. His current research uses mobile robots to study aspects of insect vision which cannot be studied by traditional physiological and simulation methods.

    F. Claire Rind received her B.Sc. in Animal Physiology at Canterbury University, New Zealand, and both an MA and Ph.D. at Cambridge University, UK. She is currently a Royal Society Research Fellow at the University of Newcastle upon Tyne. Her research interests are the neural mechanisms used by insects to detect motion and the application of these mechanisms into artificial systems.

    Paul F.M.J. Verschure, born 1962, is a group leader at the Institute of Neuroinformatics, ETH-University Zurich, Switzerland. He received both his MA and Ph.D. in Psychology. His scientific aim is to find a grounding for psychological explanations in neuroscience through the use of synthetic methods. He has pursued his research working at different institutes in the US (Neurosciences Institute and The Salk Institute, both in San Diego) and Europe (University of Amsterdam and University of Zurich). He works on biologically realistic models of perception, learning, and problem solving which are applied to robots and on the software tools which support this research in synthetic epistemology, IQR421. His work has made significant contributions to the developing field of, so-called, new artificial intelligence. In addition he applies these models in the domain of art and technology.

    1

    Supported by SPP, Swiss National Science Foundation.

    2

    Supported by a Royal Society Research Fellowship.

    View full text