PR2 Learns To Learn And Generalize
Writing by Evan Ackerman on Tuesday, 29 of December , 2009 at 3:40 am
If we ever find ourselves with our own household robots, they’re going to need to be able to learn from us, because there will be all kinds of wild and crazy stuff that we want our robots to be able to do that the robot designers won’t have thought of. So, the ability to teach robots new tasks is rather important. But even more important is teaching robots to generalize one thing or task and apply it to many different variations on that thing or task. Back in September, we wrote about an ASIMO that learned to generalize different types of objects: show it a chair, and ASIMO could figure out what other objects were chair-like.
PR2 has taken this concept of generalization a step further by applying it to actions, not just things. When you teach PR2 an action (by providing an example of the action), PR2 can understand the fundamental components of the action and apply them to different variations of the action. So like, if you teach PR2 to pour a beverage out of a can into a cup, it then doesn’t matter what kind of can or what kind of cup or where they are relative to each other or if you move them around in the middle. This is hugely, hugely important for robots to be able to do, because having to teach a robot an action for “pour from can into cup on counter” and then a separate action for “pour from can into cup on low table” isn’t just tedious, it’s one of those things that is going to keep anyone from wanting to be involved with the robot at all.
There are a lot of major obstacles to surmount when it comes to creating a robot that can be practical and useful around people who know nothing about robots… From stuff like plugging into a normal outlet to recharge to not squeezing things too hard. It’s awesome to see Willow Garage and PR2 tackling these critical objectives and finding solutions that work.
[ Willow Garage ]
Comments (3)
Category: Artificial Intelligence
- Add this post to
- Del.icio.us -
- Digg
Comment by FelipeChoque
Made Tuesday, 29 of December , 2009 at 10:37 am
Wow, that looks like the beginning of a big step further!! :D Kudos to them!
Comment by Barrett Ames
Made Tuesday, 29 of December , 2009 at 1:39 pm
I love how all of their software is online too!
Comment by Joey1058
Made Saturday, 2 of January , 2010 at 1:09 am
Nice response times from the bot!