Helping Robots Sense the World Around Them

Professor Malcolm MacIver talks about a new approach to collecting animal movement data and its significance for the robotics field.

Think about the number of times in a day your body moves in reaction to something. You hear a sound behind you and you turn your head. You see something move out of the corner of your eye and you turn to take a closer look. We're constantly noticing our surroundings and adjusting our body in reaction to what we observe.  

That type of movement control is precisely what Professor Malcolm MacIver works on in his lab at Northwestern. Earlier this fall, MacIver spoke at Columbia University about a new approach he developed with Todd Murphey, Master of Science in Robotics (MSR) program director, to quantitatively understand why animals move and position their sensory organs the way they do. The talk, titled "Tuning movement to optimize information harvesting and the transition to planning," was part of the Columbia Neuroscience Seminar series.

Afterward, he took time to talk about the significance of this new approach to the robotics field.

"Current robotic sensing technology does not do well in highly uncertain environments that animals thrive within." — Professor Malcolm MacIver

How do you describe what the topic of the talk was about?
I talked about how we, as animals, deal with changing our position in space for two different ends: one, to increase the quality of information, such as turning our head toward a sound or moving our eyes when we see something move in the periphery; and two, to be strategic about where our body is during predatory interactions, like slinking along a line of bushes to surprise attack prey or hiding without movement to avoid a predator. My lab does work on both forms of movement control. 

How does your new approach differ from other existing approaches for acquiring information?
There are surprisingly few quantitative models of how animals position or orient their sensory organs in order to acquire information, by which I mean enabling the calculation of how the position or orientation should change over time. The utility of such a quantitative approach is that these 'trajectories' of sensory organs that are predicted by such a model can then be easily compared to what is measured in live animals.

One existing approach is an algorithm called infotaxis that is meant to maximize information. The way this works is you survey all possible next actions of a sensory organ, such as move forward three centimeters and rotate by 10 degrees, and then calculate for all possible actions what the gain of information will be. You then predict that the sensory organ will take the action that maximizes information.

The approach we have developed over the past five years or so is quite different from this 'infomax' strategy. We call it energy-constrained proportional betting. The idea of this approach is that sensory organs are positioned over time so as to proportionally bet on the expected information density. Thus, while an area that has 80% of the density within it will be sampled 80% of the time, locations that have 10% of the density will be sampled 10% of the time. So long as there is non-zero probability density, the location will eventually be sampled if you go for long enough.

What is the significance of this new approach?
Current robotic sensing technology does not do well in highly uncertain environments that animals thrive within. Our new approach is a fundamentally different way of thinking about what sense organs should do. For example, across multiple animal species, it predicts sensor movements that were previously thought to arise due to noise or error. These apparently unintentional or suboptimal movements represent the animal hedging its bets by examining less probable locations for the information being sought (such as where a bear is among some dark moving shadows ahead). Our algorithm not only helps us understand how animals do it, but can prescribe sensor movements for improved performance in robots. 

What types of robots can most benefit from this energy-constrained proportional betting?Any robots or other types of machines that need to operate in naturalistic environments. 

You've previously called MSR students the most ambitious and curious master's students you've ever worked with. How do you think MSR is helping prepare students to solve complex problems and continue to push the field forward?
The MSR program does a wonderful job of matching MSR students with faculty in Northwestern’s Center for Robotics and Biosystems and elsewhere so that they have an educational experience with an exciting project outcome. These projects are typically at the forefront of technology.  

What advice would you give to a prospective student considering MSR?
Be creative in envisioning what kind of project you pursue. You might be able to interest faculty in projects that are a bit outside their comfort zone. Push yourself to grow in places where you consider yourself weak.

McCormick News Article