Avik De (@avikde): "A single robot brain that can control multiple robot bodies may need to make discrete jumps in the strategy it uses, like switching from walking to jumping. A deep neural network is not very efficient at representing this kind of output discontinuity. An architecture that can do…"

Make money doing the work you believe in

A single robot brain that can control multiple robot bodies may need to make discrete jumps in the strategy it uses, like switching from walking to jumping. A deep neural network is not very efficient at representing this kind of output discontinuity. An architecture that can do that well, or a type of “brain-body compilation” could solve this issue.

However, by the nature of deep neural networks, a discontinuous divergence in strategies is not very efficiently representable. As an example of strategy discontinuity, consider that my cat needs to run or jump up human-sized stairs, whereas a larger animal could simply walk up them. The multi-robot brain needs to do a lot of work to represent effective locomotion strategies for these variations.

Avik De

May 20

12:40 PM

Make money doing the work you believe in

Log in or sign up