Nvidia introduced Eureka – an AI agent for teaching robots to perform complex physical actions

Nvidia introduced Eureka – an AI agent for teaching robots to perform complex physical actions

In its blog, Nvidia talked about the AI ​​system Eureka, which teaches robots to perform complex physical actions. For example, with its help, a robotic arm learned pen-spinning, and a manipulator learned how to open drawers.

AI agent Eureka helps automate one of the routine tasks in the field of neural networks — the development of reinforcement learning algorithms. The study notes that, for the most part, programmers have to manually try different reward options, analyze the learning results, and then adjust the algorithm.

Eureka uses GPT-4 and generative models for writing code. The AI ​​agent accepts comments from humans and edits the algorithm. At the same time, the system understands natural language and it is not necessary to report the exact parameters that need to be corrected. For more efficient training, Eureka can analyze large samples of reward candidates, selecting the best ones.

After that, the system compiles summary statistics of training results, analyzing the effectiveness. Based on this data, new training parameters and rewards are formed. Thus, one neural network trains another, receiving instructions from the developer.

Eureka has already been tested on models of four-legged robots, bipeds, robotic arms and others. For example, the system taught the coward penspinning — fast juggling of the pen with his fingers. The study also published the results of teaching a manipulator to open a drawer, and a bipedal robot to run. All experiments were performed within Nvidia Isaac Gym.

According to research results, Eureka’s algorithms outperform programs developed by industry experts in 80% of cases. Thanks to this, it was possible to increase the overall productivity by more than 50%.

Related posts