A WAM arm learns how to flip a pancake with a lot of trial and error. Emphasis on the error.
One day robots will sell like hotcakes…especially if they can make you hotcakes. Scientists from the Italian Institute of Technology have successfully taught a Barrett WAM arm the time honored culinary skill of flipping pancakes. Petar Kormushev and Sylvain Calinon first trained the robot by actively moving its arm in a demonstration. From that point on, however, the robot learned by evaluating it’s own success and trying to improve upon it. A fake pancake was used so that it could be more easily tracked. Still, it took about 50 trials to get the basics down. The video of the project is hilarious. I just can’t get enough of watching an artificial pancake whiz about a room haphazardly. Check it out for yourself below.
The real story behind this project is not the pancakes, but the programming. We’ve seen other robots perform much better in the kitchen than this WAM. What makes Kormushev and Calinon’s attempt intriguing is that they relied on reinforcement learning – a process by which a program (like the one controlling the robot) tries to maximize its success. The robot didn’t simply repeat the example is was first given, it actively asked itself what it had done well, and what it would probably benefit from changing. The researchers used variations on this premise (most of which are beyond my programming knowledge) to help the robot quickly decide what works. 50 trials may sound like a lot (and it probably is) but the fact that the robot’s program was learning, and not simply repeating a defined set of tasks, is interesting. (This focus on learning is not a surprise when you consider that Kormushev is also working on the iCub child-like robot). In the future, these techniques may help computers/robots learn tasks with a minimum of input and help them become proficient automatically over time. Flipping pancakes today, performing heart surgery tomorrow. Just don’t be one of the first 50 patients.
[screen capture and video credit: Kormushev and Calinon]
[source: Kormushev website]
Comments
50 tries really doesn’t seem like all that many, considering I tried until I just gave up.
50 tries really doesn’t seem like all that many, considering I tried until I just gave up.
There’s a great feature of machine learning: Once a robot has acquired a skill or knowledge, it can be transferred to other robots quickly and effortlessly. It doesn’t matter if the original skill was programmed, or learned after a thousand attempts — it can be shared so that every new robot is instantly as proficient as the first.
Contrast that with humans. It may take an individual many decades to acquire proficiency in a subject or skill. And then to share that knowledge with a student will require long periods of instruction, training, and practice. And in the end, the student may still not be of the same caliber as the teacher.
Imagine having hundreds of identical robots, and having each master one useful skill (like flipping a pancake). Since their knowledge can be shared instantly, every robot will immediately become proficient in all skills. A robot master chef will also be a master carpenter, surgeon, gardener, welder etc.
This feature of artificial intelligence alone will be of trememdous significance to the future economy.
There’s a great feature of machine learning: Once a robot has acquired a skill or knowledge, it can be transferred to other robots quickly and effortlessly. It doesn’t matter if the original skill was programmed, or learned after a thousand attempts — it can be shared so that every new robot is instantly as proficient as the first.
Contrast that with humans. It may take an individual many decades to acquire proficiency in a subject or skill. And then to share that knowledge with a student will require long periods of instruction, training, and practice. And in the end, the student may still not be of the same caliber as the teacher.
Imagine having hundreds of identical robots, and having each master one useful skill (like flipping a pancake). Since their knowledge can be shared instantly, every robot will immediately become proficient in all skills. A robot master chef will also be a master carpenter, surgeon, gardener, welder etc.
This feature of artificial intelligence alone will be of trememdous significance to the future economy.
Unfortunately, to flip a real pancake, you need a lot more height that than, which is the disadvantage of learning from an inflexible model pancake.
However, that is my only criticism. One robot flipping a pancake from programming is unimpressive these days, however, one robot learning to flip a pancake relatively unaided. That _is_ impressive.
Self-optimising subroutines are becoming truly incredible these days. I just can’t wait to see what they’re like in ten or twenty years’ time.
Unfortunately, to flip a real pancake, you need a lot more height that than, which is the disadvantage of learning from an inflexible model pancake.
However, that is my only criticism. One robot flipping a pancake from programming is unimpressive these days, however, one robot learning to flip a pancake relatively unaided. That _is_ impressive.
Self-optimising subroutines are becoming truly incredible these days. I just can’t wait to see what they’re like in ten or twenty years’ time.