Nvidia has announced the development of a new AI agent called “Eureka,” driven by the GPT-4 model from OpenAI. Eureka is capable of independently teaching robots complex skills.
According to the company, “Eureka autonomously writes reward algorithms and was hand-trained for the first time to perform quick tricks like pen spinning, it has also taught robots how to open drawers and cabinets, throw and catch balls, and manipulate scissors, among around 30 tasks.”
Reinforcement learning has led to remarkable advancements over the past decade, but it still poses various challenges, such as designing rewards, which remain a trial-and-error process. Eureka represents the first step in developing new algorithms capable of integrating generative and reinforcement learning methods to tackle complex tasks.
Nvidia has also released the Eureka AI Algorithms Library, allowing individuals to experiment with it using Nvidia Isaac Gym, a physics simulation benchmark application for enhanced learning research.
Nvidia Isaac Gym is built on Nvidia Omniverse, a development platform for building three-dimensional, OpenUSD-based tools and applications.
In recent months, there has been increased discussion surrounding AI agents, particularly in April with the emergence of independent AI agents like Auto-GPT, BabyAGI, and AgentGPT.
Nvidia’s current work builds upon previous efforts, including Voyager, an AI agent designed based on GPT-4, capable of independently playing the game Minecraft.
Jeff Clune, a computer science professor at the University of British Columbia and a former researcher at OpenAI, mentioned in a recent New York Times article the efforts to transform chatbots into online customers: “This is a huge business opportunity that could be worth trillions of dollars,” he said, while acknowledging the significant positive and far-reaching societal implications.
The authors of a new research paper stated, “Eureka leverages the latest generation of zero-shot learning, code-writing, and contextual fine-tuning capabilities of advanced large language models like GPT-4 to perform evolutionary improvements on the reward function, which can be utilized for acquiring complex skills through reinforcement learning.”
Nvidia stated, “Eureka is a unique blend of large language models and accelerated simulation techniques with GPU processing units. We believe Eureka is capable of intelligent robot control and provides a new way to physically render realistic animations for artists.”
