Page 71 - Artificial Intelligence for the Internet of Everything

Intelligent Autonomous Things on the Battlefield 57

careful play. The agents will produce dramatic new behaviors that are per-
fectly suited for their tasks yet display the same adaptability that we find in
their perception systems. These creative solutions to embodiment will
accomplish another important goal: to find the inflection points in the design
and control space where complex physical interactions of body and environ-
ment lead to mechanical advantage and efficient cycles that let systems over-
come their fundamental limitations on weight and power.
Humans have developed many control schemes that allow us to bend
natural processes to our will, and machines must continue these develop-
ments, tapping into a capacity for learning and adaptation to magnify these
traditional techniques. In some cases, this means identifying pieces of our
planning and control systems that are particularly hard to model and letting
the system learn them (Richter, Vega-Brown, & Roy, 2015). In other cases
we can take a more holistic view of learning to plan and control: recent
advances in planning have allowed us to sidestep the challenge of thinking
through all of the compounding choices by substituting enumeration with
sampling and relying on the mechanisms of probability to give us a path to
convergence and optimality (Karaman & Frazzoli, 2011; Tedrake, Manches-
ter, Tobenkin, & Roberts, 2010). Intelligent systems will learn to guide their
sampling over time as a way to realize the knowledge they have built about
the world and their capabilities while still retaining the ability to be flexible
in case their model is not quite right. Modeling probability distributions can
be challenging, but learning to transform a simple distribution can accom-
plish the same goal and provide a path to learnable sampling-based control
(Lenz, Knepper, & Saxena, 2015).
These planning and control approaches are founded on the ability to
make predictions about the effect of our actions on the world, but, in a bat-
tlefield environment where conditions can deteriorate rapidly and intelli-
gent things may need to continuously adapt to damage, forming these
predictions is another task that AI must step in to perform. We can take
inspiration from studies of how the human mind develops and learns in order
to understand how agents can develop intuition about the physical world
(Tenenbaum, Kemp, Griffiths, & Goodman, 2011); this intuition can be
holistic or focused on understanding specific aspects of the world such as
rigid body motion (Byravan & Fox, 2017) or how fluids behave (Schenck
& Fox, 2016). Through observations and careful experimentation (Pinto,
Gandhi, Han, Park, & Gupta, 2016), agents can treat physics as another
application of perception algorithms that can take cause and effect data