We last looked in on Swiss Mile’s wheeled quadruped robot two years ago. In the time since, it’s made astonishing progress. First off it’s gotten a lot better at standing up–faster than I’m comfortable with, to be honest–and can now use its wheels as hands:
Listen beautiful relax classics on our Youtube channel.
Now the developers have been training the ‘bot, first virtually then in the real-world, with “curiosity-driven learning.” It’s kind of scary-sounding. Essentially they use Reinforcement Learning, like dog trainers, to program the robot to earn “rewards” for accomplishing a task through experimental behavior.
“The desired task is encoded in a single sparse reward, i.e., a reward of ‘+1’ is given if the task is achieved. Intrinsic motivation enables learning by guiding exploration toward the sparse reward signal.”
In other words, rather than program the ‘bot on how to open a door, or pick up a package and place it in a box, they simply tell the ‘bot that accomplishing those tasks is rewarding. This makes the robot “curious” as to how they can do them, and they start figuring out how to use their parts to execute the task. Check it out:
Well, I can’t see this going wrong…as long as we’ve got a good handle on who programs what is and isn’t rewarding.
Source: core77