Want a sandwich? Yeah, me too. PR2 has learned how to fetch stuff from the fridge, which is great and all, but thanks to a technique called semantic search, it can now bring you a sandwich when it's not even sure where the sandwich is.
"Semantic search" is simply the ability to make inferences about an object based on what is known about similar objects and the environment. It sounds complicated, but it's really just a computerized version of what we humans think of as "common sense." For example, if someone asks you to bring them a cup without telling you exactly where the cup is, you're probably clever enough to infer that cups can be found in drawers or cabinets or dishwashers, and that drawers and cabinets and dishwashers are all usually located in a kitchen, so you can go to the kitchen, poke around for a little bit, and find a cup. Semantic search allows robots to do the same sort of thing.
The advantage of this technique is that it gives robots the ability to infer things that it doesn't know from things that it does know, and use reason to make deductions about parts of the word that it's less familiar with. Additionally, the robot can add to its knowledge base to quickly adapt to new places and people with weird habits. So like, if you're one of those people who stores peanut butter in the bathroom, the robot can start associating peanut butter with bathrooms.
The following demo, from the University of Tokyo and Technische Universität München, puts semantic search to the test by tasking a PR2 with fetching a sandwich. The PR2 has no detailed information on sandwiches, but its database tells it that sandwiches are a type of food, and that food can be found in kitchens and restaurants, and from that, it figures out where to look:
PR2, which already knows how to get drinks and bake cookies, managed to get from a simple "bring me a sandwich" command to going and ordering a sandwich at a Subway in a separate building, all entirely autonomously. Yeah, it had a little trouble with the elevator buttons, but let's think about the big picture: using techniques like these, we're getting closer to being able to give robots vague and general commands, and having them figure out what we want really want and how to make it happen.
"Semantic Object Search in Large-scale Indoor Environments" by Manabu Saito, Haseru Chen, Kei Okada, Masayuki Inaba, Lars Kunze, and Michael Beetz from the University of Tokyo and Technische Universität München was presented last week at the IEEE International Conference on Intelligent Robots and Systems in San Francisco.