A Synthetic Nervous System Model of the Insect Optomotor Response
313
What is known about the structure of the optic lobe? There are three separate neuropils
contained within the optic lobe: the lamina, medulla, and lobula complex, which itself
consists of the lobula and the lobula plate (for a review see [3]). Each of the neuropils
contains organized columnar units corresponding to the ommatidial (i.e. lens) array in
the retina and operates on neighboring columns in each successive layer. The lamina
is stimulated by the retina and inhibits its neighboring columns to increase the contrast
and dynamic range of incoming images. The medulla appears to correlate the activity
of adjacent columns with a time delay in order to detect motion across the retina via
“elementary motion detectors” (EMDs), the precise structure of which is not known.
The lobula plate contains cells that run tangentially to the retinal columns and sum the
motion responses of the medulla across the visual field. The output of these Lobula Plate
Tangential Cells (LPTCs) encode the wide-field motion of the visual scene. Interneurons
mediate these signals to the motor neurons in the thoracic ganglia, enabling motor centers
to move the head or body in response to the wide-field motion [5].
Such interneurons that communicate wide-field visual cues with the motor networks
are critical because such cues are primarily generated by the animal’s own motion.
Therefore, minimizing the optic flow is one way that insects may stabilize their gaze or
posture. Simulating wide-field motion by displaying moving patterns that envelop the
animal have been used to evoke the “optomotor response”, wherein the animal turns
its neck [6], adjusts its posture [7], or walks along a curve [8] in an attempt to cancel
out this visual motion. Implementing such a system on board a robot may enable us to
enhance the postural stability of the robot, while providing an opportunity to model how
animals may use optic flow information to direct walking. Some robots have also used
optic flow to avoid barriers while walking [9]. As a proof of concept, we model the robot
as a single neck actuator that can rotate the “head”.
In this manuscript, we describe the robotic hardware and our neural modeling approach. We summarize the structure of the insect lobe and explain the simplifications and
assumptions we made while constructing our model. We show that the layers within our
model perform the computations observed or hypothesized to occur in the animal. We
show that the result of the visual processing is a rate-coded estimate of the speed and
direction of the background’s motion. We show that by using the output of the model to
stimulate motor neurons that actuate the neck, our robot acts as a closed-loop dynamical
neuromechanical model of the insect optomotor response. Finally, we discuss how this
system will be expanded in the future in order to incorporate more features present in
the insect optic lobe, while increasing its utility for robotic vision.
2 Methods
2.1 Robotic Hardware
Figure 1 presents the robotic hardware. The “head” consists of a Raspberry Pi with
the Camera Module, equipped with a 360° lens (Fig. 1A). The head is rotated by one
Dynamixel smart servo (Robotis, Seoul, South Korea). The “background” consists of
a paper drum, the inside of which has a stripe pattern printed onto it. This is meant to
mimic the experimental setup of studies of the insect optomotor response [8]. The head
and background are connected to the same 3D printed chassis (Fig. 1B).
313
What is known about the structure of the optic lobe? There are three separate neuropils
contained within the optic lobe: the lamina, medulla, and lobula complex, which itself
consists of the lobula and the lobula plate (for a review see [3]). Each of the neuropils
contains organized columnar units corresponding to the ommatidial (i.e. lens) array in
the retina and operates on neighboring columns in each successive layer. The lamina
is stimulated by the retina and inhibits its neighboring columns to increase the contrast
and dynamic range of incoming images. The medulla appears to correlate the activity
of adjacent columns with a time delay in order to detect motion across the retina via
“elementary motion detectors” (EMDs), the precise structure of which is not known.
The lobula plate contains cells that run tangentially to the retinal columns and sum the
motion responses of the medulla across the visual field. The output of these Lobula Plate
Tangential Cells (LPTCs) encode the wide-field motion of the visual scene. Interneurons
mediate these signals to the motor neurons in the thoracic ganglia, enabling motor centers
to move the head or body in response to the wide-field motion [5].
Such interneurons that communicate wide-field visual cues with the motor networks
are critical because such cues are primarily generated by the animal’s own motion.
Therefore, minimizing the optic flow is one way that insects may stabilize their gaze or
posture. Simulating wide-field motion by displaying moving patterns that envelop the
animal have been used to evoke the “optomotor response”, wherein the animal turns
its neck [6], adjusts its posture [7], or walks along a curve [8] in an attempt to cancel
out this visual motion. Implementing such a system on board a robot may enable us to
enhance the postural stability of the robot, while providing an opportunity to model how
animals may use optic flow information to direct walking. Some robots have also used
optic flow to avoid barriers while walking [9]. As a proof of concept, we model the robot
as a single neck actuator that can rotate the “head”.
In this manuscript, we describe the robotic hardware and our neural modeling approach. We summarize the structure of the insect lobe and explain the simplifications and
assumptions we made while constructing our model. We show that the layers within our
model perform the computations observed or hypothesized to occur in the animal. We
show that the result of the visual processing is a rate-coded estimate of the speed and
direction of the background’s motion. We show that by using the output of the model to
stimulate motor neurons that actuate the neck, our robot acts as a closed-loop dynamical
neuromechanical model of the insect optomotor response. Finally, we discuss how this
system will be expanded in the future in order to incorporate more features present in
the insect optic lobe, while increasing its utility for robotic vision.
2 Methods
2.1 Robotic Hardware
Figure 1 presents the robotic hardware. The “head” consists of a Raspberry Pi with
the Camera Module, equipped with a 360° lens (Fig. 1A). The head is rotated by one
Dynamixel smart servo (Robotis, Seoul, South Korea). The “background” consists of
a paper drum, the inside of which has a stripe pattern printed onto it. This is meant to
mimic the experimental setup of studies of the insect optomotor response [8]. The head
and background are connected to the same 3D printed chassis (Fig. 1B).