Nvidia AI agent writes ‘better reward algorithms than those of humans’

Spread the love

Nvidia Research unveiled its new AI agent Eureka on Friday. This has taught robots how to open drawers and use scissors, among other things. It has also taught a robot hand how to perform complex dexterity tricks with a pen.

The AI ​​agent uses GPT-4 and generative AI to write code for robots that apply reinforcement learning, writes Nvidia Research. The company says Eureka has, among other things, taught robotic arms and robotic hands to perform 30 different types of tasks. The AI ​​agent’s reward algorithms are 80 percent more effective than those written by human experts, the company writes. This has improved the performance of the robots by more than 50 percent. In addition, Eureka evaluates itself based on training results and looks at possible changes in the reward system. Nvidia Research has a library of the Eureka algorithms and the research published.

In a short video, Nvidia shows a simulation of a robot hand that handles a pen in different ways, which, according to Nvidia, does this ‘as well as a human’.

.fb-background-color { background: #ffffff !important; } .fb_iframe_widget_fluid_desktop iframe { width: 100% !important; }
Share