5 — Introducing MantisBot: Hexapod robot controlled by a high-fidelity, real-time neural simulation

MantisBot is a six-legged robot controlled by a realtime neural simulation. Video

An important distinction that I can’t make enough: A computational neural net is nothing at all like a biological one. So this isn’t running TensorFlow or Keras under the hood; it’s running a bunch of mini-simulations of action-potentials rushing down phospholipid membranes.

Like insects, MantisBot uses central pattern generators (CPGs) and local reflexes at each joint to control movement in a distributed fashion. This is one way real insects are able to react so quickly to environmental stimuli with so simple a brain: Most of their locomotive computations are performed in the limb itself (in ganglia), not in the brain.

The neural net is “conductance-based”, which means that it simulates neurons as physical bodies with electric potential across the membrane; using an i7 desktop processor, 775 of the mantis’ neurons are simulated faster than realtime. I mention the number because it is orders of magnitude fewer neurons than exist in a biological mantis, but is more than double the number of neurons in the well-understood connectome of C. elegans.

This conductance-based model follows

\[C_{mem} \frac{dV}{dT} = G_{mem} \cdot (E_{rest} - V) + G_{syn} \cdot (E_{syn} - V) + G_{Na} \cdot m \cdot h \cdot (E_{Na} - V)\]

where $m$ and $h$ are Na-channel activation and deactivation states. Unlike other robotic-connectome projects which use a “point-model” of a neuron, this model aims to simulate some amount of electric neurophysiology (though there are, of course, more accurate models in existence).

Using a neuromimetic model like this grants a robot some very interesting features: In particular, this robot can redistribute its body weight evenly due to its simple reflexes, even though it doesn’t have an internal body model, or even know its shape or weight: All this, for free: The robot expends no extra energy when recovering from a misstep because the reflexes are always-on.