Hey, there! Log in / Register

MIT perfecting robot that can bake cookies

Via PC World, which notes:

Mario is currently working with other students and professors to program PR2 so it can wipe down the table and put the baking sheet in the oven.

Neighborhoods: 
Topics: 
Free tagging: 


Ad:


Like the job UHub is doing? Consider a contribution. Thanks!

Comments

...by the way it just dropped empty bowls on the floor when it was done with them.

Seriously, this is pretty cool - and pretty typical of the sort of work top-tier roboticists are doing these days. Just to give folks here a bit more context - the PR2 robot is a commercially available, general purpose research and development platform. It's built so that you can program it to perform many different types of tasks. The MITers didn't build the bot, they're using it to do programming research.

Specifically, the CSAIL students are attempting to create *heuristic* ("discovery-based") programming for the robot baker. Typically, robots are manipulated with series of expliit spatial instructions - eg, "Move left arm 45 degrees clockwise, open left pincers 2.5 cm, lower left arm 30 degrees..."

This requires that everything be just-so. A robot programmed this way can do one set of tasks, and do them the same way, and cannot adapt to altered initial conditions. This is fine for an assembly line, but not so good for pretty much everything else.

A heuristically programmed robot baker, otoh, is simply presented with a table of ingredients and utensils and a recipe with typical cookbook-like instructions -= eg, "Cream butter and sugar together in large bowl. Add flour and chips and stir thoroughly..."

It "knows" what different ingredients look like, so it can figure out which bowls to use, wherever they are on the table. It "knows" how to stir things together, and what it looks like when they are sufficiently mixed. It can "see" where the baking tray is and scrape the batter in, etc. In other words, it can adapt to initial conditions and still perform a wide variety of tasks.

Quite frankly, at this stage, we're still not completely clear on how *humans* learn to do new tasks, so this sort of research has almost as much to do with cognitive science as it does with traditional engineering.

up
Voting closed 0