AI Teaching a Robot Hand to Twirl a Pen

NVIDIA showed in a video that a robot hand trained by Eureka can do cool pen tricks just as well as or even better than really skilled people.

Eureka, the smart computer, tests its learning in a special program and then looks at the data to make itself even better. It can make different robot hands that can use scissors, spin pens, and open cabinets in a computer program that acts like the real world.

Eureka’s way of learning is not just good; it’s often better than what people make. In their research paper, they showed that Eureka’s programs were better than human-made ones in more than 80% of the tasks. This means they made the robot hands work about 50% better in the computer simulations.

Reinforcement learning has done some amazing things in the last ten years, but there are still problems to solve, like figuring out what rewards to give, which can be tricky.

Anima Anandkumar, who works on AI research at NVIDIA and helped write the Eureka paper, said that Eureka is a start in making new methods that use both generating ideas and learning from rewards to solve tough problems.