We do data science.
Feb 28, 2017 • Noon van der Silk
A group from Japan have added Deep Q-Learning to a robot, fixed a set of actions: look at human, wait, wave, handshake, and taught it to pick which of those actions it should perform when it encounters various social situations. It's only sensory input is grayscale images, and depth information, and it's only goal is to maximise it's reward: it earns 0.1 units of reward for a successful handshake, -0.1 units for a socially-inept handshake attempt, and 0 for any other action. They find that the robot learns a good measure of social competence after 14 days! I think this is quite a fun application of deep learning. I think the next interesting idea is to adapt in the ability to learn new social actions, and figure out some scheme for the robot to determine it's own reward value.
Read more here!