NVIDIA Research announced today that it has developed a new artificial intelligence agent called "Eureka", which is powered by OpenAI's GPT-4 and can autonomously teach robots complex skills.

Visit the purchase page:

JD.com NVIDIA series product summary

The company said in a blog post that Eureka, which can write its own reward algorithms, is the first to train a robot hand to perform rapid pen-turning skills like humans. Eureka has also taught the robot nearly 30 tasks such as opening drawers and cabinets, throwing and catching a ball, and operating scissors.

"Reinforcement learning has achieved impressive victories over the past decade, but many challenges remain, such as reward design, which remains a trial and error process," Anima Anandkumar, senior director of artificial intelligence research at NVIDIA and author of the Eureka paper, said in a blog post. "Eureka is a first step in developing new algorithms that integrate generative and reinforcement learning methods to solve difficult tasks."

NVIDIA Research also released the Eureka artificial intelligence algorithm library for people to conduct experiments using NVIDIA AIsaacGym, a physics simulation reference application for reinforcement learning research. IsaacGym is built on NVIDIA Omniverse, a development platform for building 3D tools and applications based on the OpenUSD framework.

The craze for artificial intelligence agents has been going on for months, including the rise of autonomous artificial intelligence agents such as Auto-GPT, BabyAGI and AgentGPT in April this year.

The current work by NVIDIA Research builds on previous work, including most recently Voyager, an artificial intelligence agent built using GPT-4 that can play Minecraft autonomously. This week, the New York Times published an article about turning chatbots into online agents. Jeff Clune, a computer science professor at the University of British Columbia and a former researcher at OpenAI, said: "This is a huge business opportunity with the potential to bring trillions of dollars in revenue. This has huge upside and huge impact on society."

In a new research paper titled "Eureka: Human-level reward design by encoding large language models," the authors say, "Eureka leverages the extraordinary zero-point generation, code writing, and context improvement capabilities of state-of-the-art LLMs such as GPT-4 to evolve the reward code."

The resulting rewards can be used to acquire complex skills through reinforcement learning. "In the absence of any task-specific prompts or predefined reward templates, the reward functions generated by Eureka outperform human-designed expert rewards. In a set of 29 open source RL environments containing 10 different robot morphologies, Eureka outperformed human experts in 83% of the tasks, with an average normalized improvement of 52%."

"Eureka is a unique combination of large language models and NVIDIA's GPU-accelerated simulation technology," said Jim Fan, a senior research scientist at NVIDIA and a contributor to the project, in a blog post. "We believe Eureka will enable dexterous robot control and provide artists with a new way to create physically realistic animations."