Nvidia Research has debuted Eureka, an AI agent that autonomously teaches robots complex motor skills. Powered by OpenAI’s GPT-4, Eureka has successfully trained a robotic hand to handle a pen with the dexterity of a human — a first, according to Nvidia. Eureka has also enabled robots to do things like open drawers, manipulate scissors and toss and catch balls, along with dozens of other tasks. “Eureka is a first step toward developing new algorithms that integrate generative and reinforcement learning methods to solve hard tasks,” according to Nvidia Senior Director of AI Research Anima Anandkumar said.

Eureka accomplishes this by autonomously writing “reward algorithms” to train bots. “Reinforcement learning has enabled impressive wins over the last decade, yet many challenges still exist, such as reward design, which remains a trial-and-error process,” explains Anandkumar, who co-authored the Eureka research paper.

Developers can experiment with the Eureka algorithms using Nvidia Isaac Gym, a physics simulation reference application for reinforcement learning research,” Nvidia explains. Isaac Gym is built on the Nvidia Omniverse development platform for creating 3D tools and applications based on the OpenUSD framework.

“Hype over AI agents has been swirling for months, including with the rise of autonomous AI agents like Auto-GPT, BabyAGI and AgentGPT back in April,” writes VentureBeat, noting that Nvidia’s Eureka research “builds on previous efforts including the recent Voyager, an AI agent built with GPT-4 that can autonomously play ‘Minecraft’.”

The New York Times reports on efforts to train chatbots to be fluent online agents, quoting University of British Columbia Computer Science Professor Jeff Clune calling capably trained robots “a huge commercial opportunity” worth “potentially trillions of dollars,” though he adds that the endeavor “has a huge upside — and huge consequences — for society.”

“The work aims to bridge the gap between high-level reasoning and low-level motor control, allowing robots to learn complex tasks rapidly using massively parallel simulations that run through trials simultaneously,” Ars Technica writes of Nvidia’s Eureka moment.

Eureka also uses Nvidia’s new SteerLM, which aims “to make human annotation in AI training a thing of the past,” according to Web3 pub Decrypt, which says “SteerLM, a method that aligns AI assistants to be more helpful by training them on human feedback,” was recently open-sourced.

Also participating in Nvidia’s Eureka research are the University of Pennsylvania, Caltech and the University of Texas at Austin.