6 — Cofields: a physically inspired approach to motion coordination

Co-Fields, or “computational fields”, are a proposed multi-dimensional navigation system to enable what the authors call the upcoming “pervasive mobility” of autonomous devices, but what the rest of us just call IoT. You should click that link.

The paper is enjoyably non-technical, despite addressing some pretty technical topics (there are a few $\Sigma$s but they’re mostly harmless). The main gist is this:

To avoid making a single “agent” do too much work to compute the global state of a navigation space, broadcast a “field” (like an electric field) that the agent can use to hill-climb or gradient-descend to reach its goal.

That is, overlay real-world spaces with Co-fields so that autonomous agents simply run a basic formula to find their next action. Imagine attaching a very strong magnet to a wall of your room, and then glueing a cube of metal to your Roomba.

I don’t think the authors spent enough time addressing the issues of local minima (or maxima, depending on your point of view): It seems like this sort of field would cause a lot of agents to get stuck in minima between the overlaid fields, unless the agent was modifying or (exploratively) overriding the global Co-fields as well.

That being said, it seems like a very powerful concept to implement in Extremely Simple Agents, where, unlike the hefty desktop behind yerterday’s MantisBot, processing-power is at a premium.

I could also see it becoming very useful for cases where individual agents don’t have global information at all, or lack the resources to store/process global maps — for instance, where individual agents are simple explorers with low processing power, or where network sparsity necessitates radio-silence from agents.