The moment has come when artificial intelligence starts teaching robots. Nvidia has developed Eureka, an AI agent that can teach robots complex motor skills.

For example, Eureka taught a robotic arm to spin a pencil, that is, quickly juggle the pencil with its fingers. Of course, a virtual model of a robotic arm is trained, but this is not very important.

Nvidia’s artificial intelligence representative taught the robots nearly 30 different tasks in total; These included things like opening cabinet doors, throwing and catching balls. Some of these actions may seem very simple, but that’s only because we know how to do them automatically and without thinking.

Eureka uses the GPT-4 language model. The training took place in the physics simulation application Nvidia Isaac Gym.

Reinforcement learning has made impressive progress over the past decade, but many challenges remain, such as reward design, which remains a process of trial and error. Eureka is the first step towards developing new algorithms that combine generative learning and reinforcement learning techniques to solve complex problems

It is worth noting that the efficiency of Nvidia’s AI agent is very high. Reward programs created by Eureka, which allow robots to learn through trial and error, outperform programs written by experts on more than 80% of tasks, according to a press release. This leads to an average increase in robot efficiency of more than 50%.

The AI ​​agent uses the GPT-4 language model and generative AI to write programming code that rewards robots for reinforcement learning. It doesn’t require task-specific prompts or predetermined reward templates and easily takes people’s feedback into account to fine-tune rewards to produce results closer to the developer’s vision.

Using Isaac Gym GPU-accelerated simulations, Eureka can quickly evaluate the quality of large groups of award candidates for more efficient training. Eureka then summarizes key statistics from the training results and instructs the language model to improve the generation of reward functions. That’s why artificial intelligence is improving itself. He has trained robots of all types—quadrupeds, bipeds, quadcopters, dexterous hands robots, manipulator cobots, and more—to perform a variety of tasks.