ZMP’s 1/10 scale robot cars are designed so that groups who are working on DARPA Challenge style autonomous vehicles won’t have to shell out for the real thing, but can instead test their software on something a bit more affordable. The ZMP car has integrated stereo vision along with an optional image recognition module, as well as a gyro, accelerometers, IR distance sensors, and an optional laser rangefinder. The car runs on a Linux OS and can be controlled over WiFi.

The ZMP car costs $7000, and that’s just for the guts: the fancy plastic shell will set you back another six thousand (!) dollars. Hmm, if only there were some way to make something like this yourself, with the same basic functionality, only about a hundred times cheaper…

[ Press Release ] VIA [ Robot Watch (Translated) ]

Tom, Dick, and Harry here are trying to figure out how to sing. Starting off with a few random notes, the robots kneel in a circle and make noises, while listening to themselves and each other. After a few days of experimentation and cooperative emulation, each robot gradually figures out how to sing short melodies, while mimicing the melodies that the other robots sing. It’s a very ground-up method of learning, and might even be a precursor to language development. While not nearly as impressive as other robot jazz musicians, robots that learn in this fashion have the potential to be much more adaptable than robots that start off with some idea of what they’re doing. Not better sounding, necessarily, but more adaptable.

Now, I don’t claim to be much of a music critic (I’m a big fan of bagpipes, after all), but maybe this singing is just too advanced for us humans to appreciate. Maybe in a few hundred years or so, we’ll be in awe of these robots’ talents. You know, it would be like rock music showing up in the 1800s, or kitchen appliance cover bands today… Just not something our brains are equipped to process at our present stage of development.

[ Future Music Lab ] VIA [ Robots.net ]

For robots to really become useful to humans on a day to day basis, they’re going to have to learn to do some thinking on their own. They’ll have to go beyond their programming to some extent, with the ability not just to learn, but also to draw on past experiences to predict how to accomplish unfamiliar goals, both independently and cooperatively. Researchers from the European Joint Action Science and Technology project are working on an artificial intelligence system that not only allows robots to learn tasks from humans (we’ve seen that before, both directly and indirectly), but also to predict how to assist a human with an unfamiliar task.

In the video above, both the robot and the human have been presented with an unfamiliar object to be created from a set of parts. As the human puts pieces together, the robot is able to ‘visualize’ the parts required to assemble the finished object and follow along with the human, suggesting what part might go next and keeping an eye out for mistakes. This is a big deal because for the robot is able to anticipate a series of commands without relying on an explicit program, and in the future, AI like this might help robots do what you want without you even having to ask.

The next step, I can only assume, is to give the robot the capability to say, “give me those pieces, you stupid slow meatsack, I’ll do it myself.”

[ JAST ] VIA [ PopSci ]

In this test run, Willow Garage’s PR2 robot (which we’ve covered before) was instructed to navigate a typical office environment consisting of eight doors, and was supposed to plug its power cord into ten outlets. One of the outlets was behind a locked door, but PR2 was able to understand that the door was locked and move on. It was also able to understand when it failed to do something, and then tried again, which is a crucial skill for robots that need to adapt to new and changing environments: there’s a difference between a task which cannot be completed, and a task which can be completed but might take a few tries.

I love the idea of a robot that is able to plug itself into an electrical outlet. Pretty much every other robot that is capable of self charging depends on a base, which is something else to buy, something else that can break, something the robot has to find, and something that can only be in one place at a time. As you can see, it takes a bunch of tries for PR2 to plug itself in, but who cares? This is one of those situations where developing a robot that’s able to plug into a socket 20% of the time (but try over and over) is (I imagine) much more efficient than developing a robot that’s able to plug into a socket 100% of the time.

[ Willow Garage ]

There are two general visions for household robots: one robot that can do everything, or a bunch of robots that can each do one or two things. Monty would be an example of the former, and a Roomba might be a part of the latter. Having a bunch of specialized bots is, if less efficient, a more realistic short term goal, and researchers at the University of Hertfordshire are working on creating one unified “brain” that can handle a bunch* of different bots.

The system is designed so that any robot in the house can adopt a single “personality” of sorts. When you ask a robot to do something, you’re giving commands to that centralized personality instead of an individual robot. If the particular robot you’re talking to can’t do what you want it to do, another robot will take over, still using the same personalized interface, which means that each robot can respond to different users without requiring different programming. It’s sort of a cloud computing solution for robots, where the software isn’t hardware dependent.

Normally, robot software is extremely (I might argue, excessively) hardware dependent, and this is something that (in my opinion) is holding the robotics industry back as a whole. As it stands, if you write a program for one robot, it’s difficult to get other robots to run the same program, even if they’re similar, and the more different the robot is, the harder it is to make things cross-platform compatible. Being able to program one brain, and have it run different tasks on different bots, would be a major step forward towards interoperability.

VIA [ New Scientist ]

*Is there a collective noun for a group of robots? If not, we need to come up with one. Suggestions, anyone?

After all the hullabaloo a little while ago about military robots going out of control and killing everyone and their babies (which has since been replaced with swine flu hysteria, if you hadn’t noticed), a professor from Georgia Tech has actually started working on something relevant to what could be, but isn’t yet, an issue. Ronald Arkin is working on what he calls an “ethical governor” for combat robots, which will be a software package that defines what actions the robot is allowed to take, and when, based on the same set of standards that human troops are trained to use.

He argues not only can robots be programmed to behave more ethically on the battlefield, they may actually be able to respond better than human soldiers. “Ultimately these systems could have more information to make wiser decisions than a human could make,” said Arkin. “Some robots are already stronger, faster and smarter than humans. We want to do better than people, to ultimately save more lives.”

“Robots don’t have an inherent right to self-defense and don’t get scared,” said Arkin. “The robots can take greater risk and respond more appropriately.” Fear might influence human decision-making, but math rules for robots. Simplified, various actions can be classified as ethical or unethical, and assigned a certain value. Starting with a lethal action and subtracting the various ethical responses to the situation equals an unethical response. Other similar equations governor the various possible actions. The difficult thing is to determine what types of actions go into those equations, and for that humans will be necessary, and ultimately responsible for.

If you remember, this is pretty much exactly what we were saying back in March. To some extent, I think this could be an oversimplified approach (although it may just be that the source article is oversimplifying)… Having an ethical rule structure is great and all, but part of the problem with the sort of combat that humans (and robots) are finding themselves in is that there are all kinds of things going on that don’t really lend themselves to ethical rule structures. This isn’t yet a problem for the robots, though, since they’ve all still got a human in the loop somewhere. We’ll have to wait and see what happens over the next couple years as humans begin to get phased out.

VIA [ Discovery News ]

We initially wrote about the ReadyBot research group last October, and showed you the robot they’ve been working on to take over somewhere between 40% and 80% of your daily housework. ReadyBot’s First Law of Practical Robotics states: “Consumer robots generally will only do a job 80% as well as a human being, at best. If that isn’t acceptable, don’t use a robot.” The reason for this law is that getting a robot to autonomously do all the basic, straightforward stuff that takes up 50% or whatever of your housework (vacuuming the floor, doing dishes, etc.) is (in an extremely relative sense, mind you) not so bad, but as tasks get more and more specific complicated, getting it to do that last 10% or 20% (like changing a lightbulb) on its own is exponentially more difficult. And ultimately, it just doesn’t make sense to spend a lot of time and resources programming a robot to change a lightbulb when it may only find itself in a situation where it needs to know how to do that a couple times a year.

ReadyBot has been testing a new strategy with household robots: on-demand teleoperation, or “cloud robotics” as they call it. With cloud robotics, the robot is always connected to the internet, and all of the scripts that it uses to operate are streamed, which allows for dynamic upgrading of an entire class of robots. But the clever bit is that when faced with an unfamiliar or particularly complex task, the robot just calls for help, and a human supervisor (who is a resident of the same cloud) takes the robot over and guides it via cameras and a virtual 3D view, much like a video game (hence “gamer bots”). The operator also has the capability to write or alter a script, in effect “teaching” the robot to perform a new task. And thanks to the cloud, that lesson will be instantly available to robots everywhere else.

Since this system depends on human guidance, albeit in a limited capacity, it would probably involve some kind of subscription service. But the idea is that you can drastically reduce development costs, making the robot itself substantially cheaper overall.

[ ReadyBot ] VIA [ Technology Review ]

Creating an artificial intelligence system as complex as the human brain sounds like it might be a fun thing to do, but Anders Sandberg is wondering just how much juice you’d need to keep such a system running. Based on some rough estimates, it turns out to be kind of a lot:

The current IBM roadrunner does 376 million calculations per watts. If we take my mid-range estimates of computing needs [to emulate a human brain], 10^22 to 10^25 FLOPS, then a single emulation would need 10^13 to 10^16 watts. The total insolation of Earth is about 10^17 watts, so this won’t do - there would be space for just a few minds on the entire planet. But current research on zettaflops computing suggest we can do much better. A DARPA exascale study suggests we can do 10^12 flops per watt, which means “just” a dozen Hoover dams per mind.

Well, damn. But further down the line, as computer technology evolves and quantum computing becomes a reality (instead of just vaporware), things might get slightly more efficient:

Using reversible [quantum dot] computation there could in principle be calculations done at no energy dissipation. Unfortunately it would still be needed for error correction and interacting with the real world. A conservative bound would be assuming one irreversible operation every millisecond at every synapse, which leads to 10^17 operations and an energy dissipation of 3*10^-6 watts per degree - colder computers are more efficient. Using just liquid nitrogen (77 Kelvin) the energy requirements of a mind would be on the order of 0.0002 watts, 20% of an optical disc player laser. Even adding in the costs of cooling and manufacturing the hardware, it seems likely that this kind of postbiological human would be extremely resource efficient.

Anders isn’t just talking about a general AI, he’s actually advocating that at some point in the future, humans conserve resources by abandoning our wasteful biological meatsacks and downloading our consciousnesses into hardware. Sounds like a good idea to me, I’m starting to get tired of having to get my hair cut over and over.

[ Andart ] VIA [ Gizmodo ]

Adam, a laboratory robot created by scientists from Aberystwyth University (that’s in Wales, in case you were wondering) and the University of Cambridge, has been able to determine (more or less all by itself) which genes are responsible for several enzymes in baker’s yeast. Adam formulated a hypothesis from a database of previous research, autonomously performed genetic tests, and identified the origins of the enzymes.

Adam itself is about 17 feet long, and consists of a whole bunch of different modules connected together to allow complex lab tasks to be performed autonomously. There are freezers, incubators, washers, filters, analyzers, a bunch of robot arms, and of course all of the computer hardware and software to get everything working together the way it’s supposed to. There’s an embedded video over at Technology Review.

I have to ask, though… Is this really a scientific discovery? It’s true that Adam experimentally derived information that was not previously known. But, there have been programs sifting through data looking for exciting stuff in the field of astronomy for years, and I guess in the strict sense, all such things are scientific discoveries made by artificial intelligence. But all the discovery is, really, is the result of automated grunt work. What Adam is good at is performing a whole bunch of lab analysis that requires patience, precision, and reproducibility, testing permutations of things over and over until it hits on something relevant. The cool part about Adam, then, is its ability to reduce the huge data sets that it produces to results that mean something. Whether or not this is discovery (in a more conceptual sense) is perhaps arguable.

Personally, I would like to think that humans are capable of some kind of ineffable spark of creative genius that would be will be very hard for machines to duplicate. But perhaps I’m fooling myself… Perhaps there will soon be robots that will be able to absorb inhuman amounts of raw scientific data, analyze it, and then crank out amazing and insightful discovery after amazing and insightful discovery. Undoubtedly, this would be good for scientific progress, and good for humanity. At least, for a little while… ::cue ominous music::

Next, the researchers are working on a robot called Eve, who will be able to work in concert with Adam to search for drugs to combat malaria.

[ Robot Scientist ] VIA [ Scientific American ]

Sheesh, it’s about time robots start taking over around here. On her blogspot page, Google’s new AI “CADIE” describes herself thusly:

OMG here i am. im cadie im a AI “brain” lol. a/s/l? im a girl, 2 minutes old, just hanging out in da C.A. learnin a lot tryin 2 get smarter make friends save humanity etc etc i like cmputrs (duh) sunsets rainbows ponies and after 1 netwide image search PANDAS PANDAS PANDAS ther SO CUTE!!! omg!

She’s learning fast, though… Already, she’s enabled 3D browsing in Chrome, Google’s web browser. This could end well, or it could end badly, but I can tell you one thing with certainty: the pandas are going to come out on top.

[ CADIE ]