Most animators would agree: making a cataclysmic explosion destroy a planet is easy, but human figures and delicate interactions are hard.
That’s why engineers from The Georgia Institute of Technology and Google Brain teamed up to build a cute little AI agent — an AI algorithm embodied in a simulated world — that learned to dress itself using realistic fabric textures and physics.
The AI agent takes the form of a wobbling, cartoonish little friend with an expressionless demeanor.
During its morning routine, our little buddy punches new armholes through its shirts, gets bopped around by perturbations, dislocates its shoulder, and has an automatic gown-enrober smoosh up against its face. What a day!
Beyond a fun video, this simulation shows that AI systems can learn to interact with the physical world, or at least a realistic simulation of it, all on their own.
This is thanks to reinforcement learning, a type of AI algorithm where the agent learns to accomplish tasks by seeking out programmed rewards.
In this case, our little friend was programmed to seek out the warm satisfaction of a job well done, and we’re very proud.
READ MORE: Using machine learning to teach robots to get dressed [BoingBoing]
More on cutesy tech: You Can’t Make This Stuff Up: Amazon Warehouse Robots Slipped On Popcorn Butter