A Robot Responds To Its World

Matt Bunting, engineering senior at University of Arizona, has received rock-star attention for his creation of a new version of a hexapod crawling robot which went viral on YouTube. He responds with disciplined humility,

[The attention] is definitely very new to me, but it is kind of nice in a sense. I did put in a lot of time working on this project — way to many hours compared to everyone else on this one project, but I was happy building it! You think it’s cool, but you wonder if everyone else will.

When he talks hexapod though, it’s a little difficult to pin down exactly which one:  The virtual CAD image existing as a dream in his mind’s eye since he was in middle school and started collecting parts? A first, crude artifact created incorporating his own inverse kinematic algorithms that allowed it to orient and position its legs in three dimensions?  Copies he has sent around the United States and overseas in response to wild demand for demonstrations at engineering and technology conventions? Or perhaps the next model he is already envisioning for summer 2010 that will infuse a genuine neural network with artificial neurons that communicate with one another?

In fact, the hexapod represents a constantly evolving continuum with near-daily upgrades in optical systems (a web camera is mounted at the front), leg articulation, and most intriguingly, coding that permits his robot to learn and optimize its own responses to the world around it. This is further aided by sensors in its feet. For now, it will serve to focus on a current hexapod, or spider. This one is unnamed but emblazoned with the proud capital A for the University of Arizona Robotics and Neural Systems Lab where it was incubated, Matt’s name, and the all-important Intel logo across the computer module that sits atop this promenading automaton as if sponsored for some surreal overland grand prix.

Far from following a mantra of constant improvement based solely in market motivations, this constant flux in form, capacity, and algorithms crystallizes the genuine essence of Matt’s hexapod as a set of passionate, evolving inquiries into the cognitive science, biology, and machine science. Employing intended double-meaning, this is an articulating vehicle for learning about the very nature of learning. Matt explains,

The interesting thing is with an optic-neural system, is that you can explore how a real neural system works. You come up with a theory, try to implement it, then “see” backwards. In the research on robotics, you can actually learn more about biology. You really go back and forth. I’m especially interested in tying vision to leg locomotion — I’m using Open CV (Open Source Computer Vision). I am comparing images to see how the machine is learning to walk and doing object recognition.

Meet the Spider

Activated with a standard game-controller connected by common plastic-coated wires, the robot springs to life with animistic motions that eerily evoke the spider despite being short two appendages. The six it has have three degrees of freedom. With the input from the controller, it sidles, pitches and leans, rises up substantially, skitters across the floor, and experiments with objects and levels presented to its path.

The fact that the Intel processor and chipset sit prominently atop the spider has drawn substantial attention from Intel’s Embedded Design Center. Keen to demonstrate that processors can be used flexibly for a wide range of non-computer applications, Intel has purchased a number of Matt’s hexapods and is using them for demonstrations at various venues.

Principles of Design and Production

Matt relates a convergence of opportunities that arose last year, including access to a 3D printer,

When I was taking a Cognitive Robotics course I wanted to apply ideas of cognition into an actual machine as a project — a perfect excuse to spend time and money on a hexapod. I kept eying this big white box in the lab, the 3D printer, because I wanted to give this thing (the ugly mixture of assembled parts of his high school pod) a makeover. With the 3D printer you can go all out with AutoCAD programs without worrying about feasibility (as with conventional fabrication processes). You can make any part relatively inexpensively. The printer is about $20 grand, but all the plastics for the legs in one hexapod are about $15. I’ve designed a mechanical assembly and had it the same day.

In addition to this critical short-cut to build 3D composite components, auto-analytical capacities were an essential goal for Matt:

I included something called Q Learning, which is a method of reinforcement learning — the idea is that it would do whatever it could, kind of experiment with itself to learn how to walk straight forward. You could have this machine learn how to do something — just one of the cognitive aspects.

Matt’s robot is capable of analyzing its behavior in relation to its environment, and producing the most efficient walking mechanics for the context, even if an appendage is damaged or lost.

Having difficulty viewing the video? Access the video on YouTube.

Robots as Assistive Technology?

It is becoming reasonably common to see footage of a variety of wheeled or legged robotic devices being prepared for navigation, exploration of harsh or dangerous environments on and off the Earth, or to deliver small payloads or carry instruments for data collection.

While applications of robotics for assistive roles have not been well explored, it is clear that the consideration of technology transfer opportunities is essential for niche markets like that occupied by the assistive technology industry. Movements in both directions, and hybridization, have profound implications for the market, cost efficiency, design and production, adoption rates, adaptation, new waves of innovation, and cultivation of new market space.

Matt’s background and interests currently lie outside of disability issues, but he believes the research he and others are doing in this field can benefit people medically. Leading up to one example, he starts,

I took a class a year ago called “Reverse Engineering the Fly.” It was the study of the visio-motor system, vision all the way to processing and back to how the fly then moves around in the world. Every day I think about the ideas that were presented there, especially things like optical flow. I always tend to think how I do something and how I react, how I sense something, and how could I implement that.

Continuing his train of thought, he explains that in his next iteration of the hexapod with a neural network,

I’m thinking of using Central Pattern Generators (CPG’s), basically two neurons inhibiting each other. They fire, resulting in an oscillatory behavior. From this, you can build a walking gait, and you can modify this so that it can learn how to accomplish certain things.

Citing related research in which scientists are sending extrinsic impulses to the spines of deceased research animals to  “get this walking motion,” he posits, finally, that it may be possible, partly through the incorporation of study with CPG’s and robotics, to develop therapies for “people who are paralyzed: even though you have lost the connection between the brain and lower down the spine, we could study more how this works and possibly reconnect everything.” 

He notes also that people have to learn to control electronic prostheses and adds,

Nerves in a human send signals in unknown ways, especially higher up in the spine. If a system such as a prosthetic arm could read many nerves, whether or not they are the correct nerves for arm movement, then the system could learn how to interpret the signals for proper movement. For example, is a person tries to bend their elbow, then you tell the arm that it should be bending its elbow, and the person acts how they normally would to move their elbow. The system then records the nerve signals, and this process is basically repeated for all natural movement. If a system is advanced enough, then the prosthetic could vastly increase the learning curve to operate such a device.

There is current interest in AT devices that provide video, audio, graphical, and text-based support to persons with a range of cognitive disabilities to allow them to function with daily living and basic employment skills.

It’s not too far-fetched to conceive that hexapods like Matt’s could be a mobile platform fused with programmed support information, or possibly even enough artificial intelligence to analyze client actions and offer basic corrective feedback. This could be especially beneficial for people who simultaneously have needs traditionally met by a service animal. A service pod that would not perish of natural causes and could quickly be sent reprogramming packages for new needs or users could offer substantial advantages.

Neurons on the Horizon

Digital programming, as Matt points out, has much to recommend:

The nice thing about how you program digitally is what if I accidentally design something wrong, I didn’t connect something right or if I have a low pass filter on the wrong time constant, all I have to do is change a little variable,  I don’t need to redesign my circuits.

Yet the biological domain Matt is tapping into functions largely on analog levels and principles that have their own benefits. Signals and even time function differently. Interconnecting analog and digital technologies poses substantial challenges. He continues,

Neurons work really well in the analog world, but this is a digital system. You would have to port everything to this digital program, analyze how neuron systems work and then code and program the responses, and connect them all correctly. It would be absolutely wonderful to have an analog computer where I could just connect the wires, but reconfiguring and building a lot of little analog devices is really tough.

But for this fresh-faced senior with a clearly “logarithmic” future, there’s time!