The sense of surrounding space is critical to the independence of an agent, either a biological one like mammals or an artificial one like robots. Even though SLAM has been studied for several decades in robotics, there is still no universal or context-independent spatial matric. A recent paper in this year shows that grid cells emerge with the training of recurrent neural network, and those cells help to do vector-based navigation in challenging environments.

Emergence of grid cells

The first part of the paper describes the emergence of grid cells from the supervised learning of a deep recurrent network. With the input of translational and angular velocities, and with the output of head direction and place cell activities, some neurons appear to have the grid representation.

From the description of the input, the aforementioned neural architecture basically performs dead-reckoning. Thus, we may expect that without the input of the group-truth location and the noise in those input signals, the grid representation may deteriorates along with time.

Goal-directed navigation

The following part of the paper uses deep reinforcement learning to train an agent to find the hidden goal. In this experiment, no group truth location is provided, and the the input velocities are noisy.

Another advantages of biological agent include the training time is relatively short, and the integration of short and long range space. I am looking forward to further investigation that can revolutionize the SLAM techniques in robotics.

  • image from DeepMind

Next Post Previous Post