Simulated zebrafish and a fish robot equipped with vision reveal how the body shapes brain circuits

Simulated zebrafish and a fish robot equipped with vision reveal how the body shapes brain circuits
4 min read

Researchers from EPFL, Duke University, and Instituto Superior Técnico have developed a realistic simulation of a zebrafish larva, as well as a robot inspired by this animal capable of swimming against a current using vision alone. Their work, published in Science Robotics, shows how an animal’s body and neural circuits work together to produce behavior.

To hold its position in a river, a fish must know how fast to swim and how to orient itself to compensate for the current. Most species use vision to detect optical flow, i.e., the scrolling of the scenery, and their brain transforms these signals into compensatory movements. Neuroscientists call this stabilizing reflex the optomotor response (OMR). The retina detects the direction of flow, neurons in the pretectum, a brain region, interpret it, and spinal nerves command muscle contraction.

The problem lies in the difficulty of modifying a living brain to test the role of a connection. While imaging now allows recording and even manipulating neurons during behavior, rewriting the connections of a complex animal remains nearly impossible.

A wiring diagram from the fish brain

The starting point is in the Naumann lab at Duke, where detailed behavioral studies and whole-brain calcium imaging of larvae exposed to visual stimuli mimicking the optical flow of a riverbed, combined with circuit modeling, enabled the creation of a brain-wide wiring diagram of the OMR pathways.

Based on this model and other data on neuronal processing in the retina and spinal cord of vertebrates, Dr. Xiangxiao Liu and Luca Zunino, from the EPFL team, developed simZFish, a neuromechatronic simulation. It replicates the larva’s body, with its eyes and fins, and uses this experimental wiring diagram as its brain.

simZFish, a brain that can be taken apart

Built in the Webots physics simulator, simZFish represents a six-day-old larva at 1:1 scale: a 4 mm body weighing 0.3 mg, seven segments, six simulated motors, a head with two lateral cameras, and realistic hydrodynamics. Its artificial brain reproduces the entire neural circuit, from pixels captured by the retina to muscle activation.

Researchers can modify any element of the model, remove or add neurons, or change the eye lens, then observe the effects immediately. In a live animal, the authors note, one can only record correlations of activity when the fish exhibits the studied behavior, without being able to rule out other ongoing processing. The model thus makes components identifiable and allows pinpointing the “minimal essential elements” of the OMR.

The body is part of the computation

By placing the cameras on the sides of the head, the researchers found that the optical flow produced by the current creates contradictory vortices from one visual field to the other, a case close to the classic “aperture problem.” When the entire retina fed the circuit, these signals canceled out and the OMR behavior disappeared. It returned as soon as the input was limited to the lower and posterior part of the visual field.

This area precisely corresponds to the one that most strongly drives the OMR in the real fish. According to the authors, the arrangement of the lateral eyes explains why the circuit is wired this way: it captures the most useful information with the fewest connections, an efficient solution that evolution also appears to have favored.

A prediction that leads back to the microscope

The simulation also revealed the limits of the initial model. When each eye received motion in the opposite direction, a shear stimulus never tested before, simZFish 1.0 rotated much more than expected. Shown to real larvae, the same stimulus caused almost no rotation, according to Dr. Matthew Loring, the team’s lead neurobiologist.

The Duke researchers then recorded tens of thousands of neurons using two-photon volumetric calcium imaging. They identified several previously undescribed subtypes of binocular neurons in the pretectum. Integrated into the model, they yielded simZFish 2.0, which much better reproduces the behavior of the real fish, including in response to the shear stimulus. Auke Ijspeert, who heads the Biorobotics Laboratory (BioRob) at EPFL, summarizes: “simZFish narrows the search field, animal experiments verify the predictions, and together they complete the neural circuit puzzle.”

The ZBot robot in the river

The circuit was then transferred to ZBot, an 80-centimeter robot weighing 2.7 kilograms, equipped with two real cameras, six tail motors, and a Raspberry Pi running the same neural network. In the Chamberonne river near Lausanne, the robot stayed in the field of view of a drone camera for 58 seconds with its visual circuit activated, compared to about 37 seconds when its “eyes” were deactivated and 20 seconds when it was drifting with motors off.

These trials suggest that a fish could, in principle, fight a current using visual circuits alone, which contradicts the hypothesis that rheotaxis requires the lateral line, that network of sensory organs along the body. For robotics, the approach avoids adding hardware: drones typically use a downward-facing camera, whereas the method exploits the lateral cameras already present. The simulator models and robot designs are open source.

S’abonner
Notification pour
guest

0 Commentaires
Le plus ancien
Le plus récent Le plus populaire
0
Nous aimerions avoir votre avis, veuillez laisser un commentaire.x