ASIMO Learns To Recognize Types Of Objects

Writing by Evan Ackerman on Monday, 14 of September , 2009 at 1:00 am

Man, that Cracked video just keeps giving up gems. Turns out that the bit about ASIMO being able to identify objects by class came from a BBC show called James May’s Big Ideas… being a long-time Top Gear fan, I was able to immediately recognize his shaggy locks in the 0.5 second that they were visible in the video. Anyway, I’d like to reiterate what May says in the beginning of the segment, because it’s something that we tend to point out a lot around here:

“Robots and computers are very very good at things we find very difficult, such as long division and VAT returns. But they’re very very bad at things we find extremely easy and instinctive, such as walking, talking, and seeing.”

Lots more, after the jump.

Robots are binary creatures, and as such, they have particular difficulty with anything that resides in an analog reference frame, or specifically, things that are hard to define with a set of rules. This is what makes getting a robot to talk is such a challenge: there are far too many variables for the robot to consider, and there’s no way to generate a simple set of instructions for the robot to follow.

This is why teaching robots how to learn is such a big deal, and this video explores that concept in two different ways. Firstly, ASIMO is learning how to teach itself in the same way that kids do: it pays attention to things that are waved in front of it, and listens when you describe those things. More importantly, however, ASIMO has been programmed to make inferences about unfamiliar objects based on experience and knowledge it has with familiar ones. Once it has been shown a chair, it can identify a stool or other chair-like objects. And to get back to my initial point, what this means, in a practical sense, is that the number of rules that ASIMO needs to be programmed with to be functional in this particular context is reduced from infinity to one.

Okay, yeah, that’s a bit of an exaggeration… But, consider: if ASIMO can’t understand how to extrapolate on what a “chair” is and therefore needs to be explicitly taught what objects are “chairs,” you’d have to provide it with instructions on just about any sort of chair it might encounter for it to be reliably functional around chairs. And that’s a lot of instructions.

It’s hard to overestimate how critical this type of learning is to getting functional robots into our lives. Nobody wants to have to explicitly program minutiae into our robots; we want to be able to give robots general instructions in the same way that we give people general instructions, with the confidence that those instructions will be interpreted and obeyed with a reasonable amount of intelligence and discretion. Teaching robots to learn like we do, with the ability to reason and make judgments like we do, is an important first step, and it’s reassuring to see that one robot, at least, has it sort of figured out.

Incidentally, I’d just like to mention how human this particular incarnation of ASIMO appeared while executing its learning behavior. It behaved much like a curious child would, taking steps forward and reaching for things. But then, I’m a sucker for anthropomorphizing humanoids, so maybe it’s just me.

Comments (8)

Category: Androids, Artificial Intelligence, Research

• Add this post to
• Del.icio.us -
• Digg

Name (required)

Mail (will not be published) (required)

Website

XHTML: You can use these tags: <a href="" title=""> <abbr title=""> <acronym title=""> <b> <blockquote cite=""> <cite> <code> <del datetime=""> <em> <i> <q cite=""> <strike> <strong>