Chess

This chess set may be the most awesome thing you see all day. It’s pretty remarkable, but at the same time well within reach of a dedicated group of students armed with some good reference books. (... and $30k worth of legos and computers) It would not surprise me if they said that building the robots took more time than programming them.

The biggest surprise to me, actually, is that they were able to use Bluetooth to control the robots. My experience with Bluetooth in Lego Mindstorms was mostly negative: short range, few control channels. My guess is that they must have the robots listening on shared channels, and prefacing commands with some kind of ID string: that would also explain why so many of the movements are sequential.

Post-match battery charging must be a royal pain in the ass. I also wonder what kind of corrections are needed during a long match: Sensor error and actuator slip accumulate, so that over time, as a robot moves its internal position can get wildly out of sync with its actual position and orientation. With a more sophisticated sensor suite it’s not a trivial task. It must be very difficult with the equipment they appear to have.

Anyway: my hat is off to Team Hassenplug. That’s pretty damn cool.

Your Chef for the Evening Is Rebooting

I was just reading a NY Times article about robot cooks. With Japan, food, and robots, this is an article written just for me. And hey: a pan-handling robot that drinks beer! What’s not to like?

They’re all very cool, representing remarkable technical skill, and probably many long nights in the lab. But I think the reporter is missing the point of those humanoid chef-robots, judging by the juxtaposition of those with the work of the CMU Human-Robot Interaction team.

Allow me to explain in a roundabout way: There are already machines that make ramen (maybe not ones that also have knife-fights, but bear with me here) and otherwise perform many of the other tasks here. For many individual tasks, the use of a humanoid robot or arm robot represents a lack of imagination: the mental agility to imagine how a task would be performed with an unrestricted body type often comes up with far more ingenious and efficient ways of doing it.

In fact, I’d say that very few tasks really require humanoid robots. I can’t think of any off-hand. For any individual task, there will almost always be a better form. But this is not to say that it is a bad idea to develop humanoid robots, far from it. The promise of a humanoid robot, and ultimately the (proper?) motivating factor behind many of these prototypes is the same as the promise of an iPhone or something of that ilk: A flexible device that seamlessly becomes one of any number of other single-purpose devices. This is distinct from a personal computer in some important ways, but right now the primary importance is of *doing* one thing at a time (whatever else it may be *thinking*, if you want to put it that way). By adding more cooking jobs to the general robotic repertoire, they’re converging on a suite of tasks for which the humanoid form probably is better-suited.

Microsoft Robotics also kind of gets this, I think. They ought to, anyway, as it is an analogue to Microsoft’s original strategy and success: of standardizing the slow part (the hardware) to focus on doing as much as possible with the fast part (the software). A humanoid robot (or more simply a single arm) can mechanically do just about any task they might desire (if inefficiently), so if we standardize on that ideal, the software and the logic can take a more central place. It represents a sort of design convergence: when you try to combine tasks into the simplest possible hardware, the more human tasks you add, the more human the hardware is going to look.

As for the people focused on human-robot interaction, there are interesting research questions there, and good science being done. But that research, to my mind, is not so much robotics research as it is human research with some very difficult test equipment: kind of like when zoologists design puppets that baby animals will feed from. (I really wish I could find a copy of a particular Calvin and Hobbes to link to here. It’s in “There’s Treasure Everywhere”, page 148)

Anyway, that’s my two cents on the subject. (And keep in mind that I’ve never actually done humanoid-robotic research, having focused entirely on rover-types, so I could be totally off-base)

Oh! If after reading that article you’re wondering what okonomiyaki is, by the way, it’s often referred to as a cabbage pancake or pizza. It’s... neither, really, or maybe both. I’m familiar with Osaka-style okonomiyaki, but as anyone will tell you, it can vary wildly, especially by region. For me, the little okonomiyaki-ya outside my dorm at Gaidai is the only true form: You take a batter of flour, potato starch, egg, and shredded cabbage, and spread it out on a hibachi table for some high heat, usually spread on top of some kind of meat filling like bacon or shrimp. Flip it once (so the ‘filling’ is now on top), finish cooking, then top it. The traditional toppings, to my mind, are a thick sugary sauce (like yakisoba sauce or BBQ sauce), Japanese mayonnaise, bonito flakes, and powdered seaweed. It’s... tastier than it sounds? I like it, anyway.

One final thing: I’m trying out new blog software -- MacJournal 5 from the most recent MacHeist. The interface isn’t too bad, and I do like the ability to keep separate journals in the same interface, plus locally-organized stuff: one of my big complaints for my current writing software is that it’s difficult to manage multiple projects.

Tagging seems to be more difficult compared to the web form, which autocompletes and shows me a list of tags I’ve already used.

Why Didn't I Think of That?

It's not often that I see something in the news or online that I want to smack my forehead and regret not thinking of it. This is one of those cases: moon exploration via telepresence. The basic idea is to send a slew of humanoid robots to the moon, each equipped with a variety of sensors and actuators to roughly correspond to human operators down on Earth.

The trouble is, the article doesn't give due weight to the three-second delay. Three seconds is huge, as anyone who played first-person shooters online in the 90s knows. Lag is a major factor in error, and three seconds takes a human operator well out of the range of reacting naturally to events. In the paper linked, they only go as far as 225 ms lag time, an order of magnitude faster than is being talked about here. I have no doubt that some operators will become very skilled at this delay (though I really wouldn't want to share a road with them driving home after a long session!) but the learning curve will be awfully steep, and potentially expensive.

Part of the solution to that problem will be extensive simulation. Simulators can be built right now that would get the broad strokes right in terms of gravity. The lunar regolith composition will be tough to get the feel right, but hey, they have access to people who've been there and can offer pointers.

To me, though, the answer is to scratch the humanoid part of the telepresence plan. The more the operator expects the robot to act like a human, the more frustrating the experience will be. Locomotion would be the first target -- even in the low gravity, it would be far too easy to trip and fall. Walking would quickly become a chore anyway. A wheeled or treaded robot would be my first choice, but would not be terribly mobile, and those mountains would be awful temptations. I would go for a centaur-like robot, maybe built on a Big Dog chassis: an internal controller would take care of staying upright. The feet would have to be modified, but that's do-able.

The user-controlled manipulators could sit on top of that, taking care to keep the center of gravity low. I would personally prefer (at least for the first few attempts) a modal system where the operator can either move the robot or move the manipulators, not both at once. Part of this is so that the operator's time isn't divided, but also to allow centering and balancing routines to ensure that the robot doesn't fall over.

As for getting a bunch of robots to the moon in the first place? If only there were a commercial competition devoted to getting robots onto the surface of the moon cheaply...

My God, It's Full of ... Puns

Just when I thought that there was no hope for artificial intelligence, my mind is blown by this. (h/t to PhysOrg) That's right: a computer that formulates and tells jokes, all based on that apex of humor, the pun. It is a little scattershot in its basic approach based on using Princeton's WordNet as a relational database / expert system (using the phonetic form to roughly judge whether two words sound similar. It seems to come up with some interesting results -- they remind me in some ways of the kinds of jokes five-year-olds tell. Some samples:

What do you call a cross between an emporium and a success?
A department score

What do you get when you cross a choice with a meal?
A pick-nic

And (possibly the best of the bunch)
Why is a bronzed handle different from a fringe benefit that is lordly ?
One is a tanned grip, the other is a grand tip

Oh my sides, they split.
(More here)

So, now we know how the inevitable robot armies intend to kill us all. I didn't think they'd be this cruel or ruthless, but I must now stop to wipe away a tear of pride. *wipe*

The software will eventually be available for download. I think that (with permission) this would actually be worth learning how to write iPhone apps.

What an awesome web we weave

The last thing that you want your robot army to do, when asked "Are you exterminating the human race?" is answer "Yes." Public relations fiasco right there. You can't possibly go on Oprah often enough to make up for that one.

The solution? Make sure your robots know how to lie. Fortunately, there have been great strides in this very field recently. Some folks in Lausanne, Switzerland have been working in evolutionary robotics: they developed a pretty nice platform (the S-bots, shown in that link towing a terrifically tolerant tot) that have a number of general-use actuators and sensors. At the same time, they developed a decent computer simulation of those robots so that they could try out control schemes -- including multiple iterations of an evolved controller.

Note that they have both light rings (three colors of LEDs) and light sensors. The robots are each controlled by a neural network, the inputs for which include input from those sensors, and the outputs drive motors and other actuators. The weights of each node are determined by elements on a single long string, which is treated as genetic material for a genetic algorithm. They run the robots with a randomized genome, and those robots that "survive" (get to a power source) have their "genes" passed on to the next generation (that is, the next iteration of the experiment) Neural networks are a nice handy way to bridge the gap between genetic algorithms and programmed behavior, actually. Anyone who's interested in playing around with these techniques would not do badly to start off with something like this.

Anyway, I haven't read the paper too carefully, but it sounds like the deception involves the fact that the power source (the "food") has one colored light, and the robots can change color to that same color when they are there (thus helping the other robots with different genes) or they can turn off their lights or turn a different color to give themselves a larger relative advantage! Of course, robots that are attracted to both color lights would gain a further advantage, thus increasing the dominance of that particular strain: a bunch of them with the same set of mutations could effectively hide the power source from other robots.

There's more coverage, with more videos at Singularity Hub. If you're interested in the bots themselves, the older swarm-bots have a page here, and the newer version is over here. The papers there are pretty approachable for technically-minded folks not in robotics, I think, but are not quite to the level of general consumption. Good stuff.

I guess the only remaining question, then, is that, having now learned to lie, will robots ever learn ... to love?

The answer is no.

Real Robot Facts (also fun)

Something like 50% of the people who come to this blog are searching for "robot facts" and getting a joke post from earlier. That's... kind of embarrassing, actually. Here's some more useful stuff if you want to learn about the state of robotics:

Fact #1: Japan has the coolest robots in the world. Dr. Hirose is the guy to look at, his lab produces robot snakes that are just awesome. I first came upon his work at a symposium for search and rescue robotics at NIST, where his students had a 3m-long remote-controlled robot snake for use in searching deep rubble piles. You could spend all day looking at the videos on his page and still not see it all -- if your time is a little more limited, at least watch the swimming snake bot video.

Fact #2: One of the smallest robots in the world was built by a classmate of mine at Dartmouth. They rely on having a specially-designed platform that delivers voltage levels that not only power but steer them using alternating strips of conductive material. By raising and lowering the relative voltages they can not only power one tiny little robot, they can cause its arm to stick to the surface, making it pivot (A high voltage causes it to snap downward, making use of physical forces normally too small to notice, then another reversed voltage snaps the arm back so that the robot can continue on)

Fact #3: People have more senses than they think. Proprioception in particular is interesting: it is basically the sense of where your body is with respect to yourself, and lets you do things like touch your nose with your eyes closed. Like your other senses, it can be dulled by alcohol, hence that legendary road-size sobriety test. This can be devilishly difficult for a robot -- the more limbs, the harder. Trying to calculate where the head of that snake robot is from the tail would be a formidable task if you're relying on its internal sensors alone! Sensors, like human senses, have errors that accumulate.
Related to that is the concept of dead reckoning. Try this experiment: stand with your toes to a wall, and shut your eyes. Take two steps backward, then two steps forward. You can probably manage to hot give yourself a bloody nose with two steps. Now try it with five steps (you might want a pillow in front of your face or a friend you can trust) This is a good example of the accumulation of error, and is why your Roomba vacuum cleaner can't just retrace its steps to get back to its base.

Fact #4: Because robots are being proposed for wide use in search & rescue, extraplanetary exploration, and simple patrols, there's a great deal of interest in a subject called SLAM: Simultaneous Localization and Mapping. Basically, this is the art of turning a robot into a cartographer. SLAM requires using all of a robot's sensors to figure out its pose as precisely as possible, "pose" meaning the location, orientation, and position of every part of a robot with respect to itself, to its surroundings, and (in a very interesting case) other robots on its "team". Sebastian Thrun wrote a lengthy survey paper on the subject that I still refer to from time to time.

I'd be happy to expand on these if there's interest.
(Edit: A recent post on the development of deceit in swarm robotics might be of interest.)

Fun Robot Facts

(There are real robot facts available here. This page is just lame jokes)

The movie "The Terminator" is actually a modern day remake of "Old Yeller".

Speaking of movies, did you know that the original ending to "The Matrix" trilogy was just two hours of robots beating up Keanu Reeves, with no dialog? In retrospect, they should have stuck with that.