Event:

Can a robot be made to navigate using artificial neural networks, and would the brain’s neural representations of space be useful in such a system? These are questions of interest to the field of neurorobotics, which – at the intersection between neuroscience, robotics and artificial intelligence – seeks to investigate neural networks in a robotic setting.

Grid cells are an interesting cell type in this regard, as these neurons are thought to constitute a general coordinate system of space in the brain. They are believed to participate in path integration, which is also an important process for robots to be able to keep track of their current position. Grid cells are thus a natural candidate for a spatial coordinate system to use in a neurally controlled robot.

For a robot to make full use of this coordinate system, however, the system should be able to compare different sets of coordinates in order to extract the direction of movement between them. In this way, the robot would be able to move towards arbitrary goal locations merely by knowing their respective “grid cell coordinates”.

In this talk we will see how this problem – of decoding grid cell signals into movement directions – can be solved by a biologically plausible neural network. In order to compensate for the repeating nature of the hexagonal grid cell pattern, the system integrates information from multiple grid modules of increasing scale. The system is able to rapidly calculate the required direction of movement, providing a promising building block for a future hippocampus-inspired neural robot navigation system.