Pieter Abbiel is obsessed with creating autonomous robots that can do the things that humans don't want to do. He found inspiration from a robot that swept, mopped and dusted at a friend's house. This robot, however, was programmed to do each motion of each task by its owner. Pieter envisioned a robot that could learn to do these types of tasks autonomously.
In Pieter's SolveForX talk, Autonomic Robots, he discusses his work at UC Berkeley and shows three examples of his learning robots. The talk was part of the EmTech 2011 conference, where Abbiel was crowned the winner of TR35 as one of the best innovators under 35 by MIT's Technology Review.
Pieter started with inputs and outputs. Robots see the world as a series of pixels, and then perform a series of motor commands. After the inputs the information is analyzed and the robot decides to act.
https://www.solveforx.com/moonshots/autonomic-robots
Apprenticeship learning is Abbiel's teaching method for the robots. He demonstrates a task, and then find the best modeling of the task that the robot can do, cycling through an iterative process until autonomous execution is achieved.
The first example in the video is an autonomous helicopter. The copter performs several tricks that are beyond the ability of normal humans to control. Most impressive is the backwards loop maneuver called the Hurricane.
Next a robot sorts socks. Abbiel explains that the robot was shown several hundred patterns of socks to learn colors, shapes and what a sock looked like inside out. The robot now internalizes the data and when presented with new socks can sort and ball the socks.
Finally a robot is shown sorting towels. The process looks completely unnatural as the bot takes in information about each towel and manipulates itself around the corners. Towels are folded and then neatly placed in precise piles.
This talk is very old by robotic standards but the work that Pieter has done is amazing. It's interesting for me to see the robots move - several of these motions require the robots to hold their pincers in place while the arm sections of the robot move around them. Abbiel is still doing great work with Berkeley Robotics and I'm excited to see what he does next.