LogInThey can complete other complex skills requiring high dexterity.
Can you quickly spin a pen in your hand? How long did it take you to learn this trick? It probably took NVIDIA longer to teach a robot this complex task but it succeeded, and it's a huge achievement for the industry.
NVIDIA Research has presented Eureka, an AI agent that uses a large language model (LLM) to teach robots skills that require high dexterity. It was created by the same team that made Voyager – an autonomous system that uses an LLM to teach itself how to play Minecraft in context.
Image credit: NVIDIA
Naturally, Eureka is useful not only for pen-spinning tricks, it has also taught robots to open drawers and cabinets, toss and catch balls, manipulate scissors, and more.
Powered by GPT-4, the agent generates a reward program, which enables trial-and-error learning for robots and stirs them in the right direction, and it, according to NVIDIA, "outperforms expert human-written ones on more than 80% of tasks" and leads to an average performance improvement of more than 50%.
Image credit: NVIDIA
To make it work, Eureka uses GPT-4 to write some code that rewards robots for reinforcement learning and incorporates human feedback to modify its rewards for more accurate results. The best part is that it doesn't need task-specific prompting or predefined reward templates, so it can potentially be used for any task.
Eureka constructs a summary of the key stats from the training results and asks the LLM to improve the reward functions, making the AI a self-improving system. It's flexible and works with different kinds of robots: quadruped, bipedal, quadrotor, dexterous hands, cobot arms, and others.
“Eureka is a unique combination of large language models and NVIDIA GPU-accelerated simulation technologies,” said Linxi “Jim” Fan, senior research scientist at NVIDIA, who’s one of the project’s contributors. “We believe that Eureka will enable dexterous robot control and provide a new way to produce physically realistic animations for artists.”
Image credit: NVIDIA
NVIDIA calls Eureka a breakthrough, and it's hard to argue, even though it has a lot to learn still: the possibilities it opens are amazing if you put your mind to it.
Find out more about the agent on NVIDIA's site, read the paper here, and join our 80 Level Talent platform and our Telegram channel, follow us on Instagram, Twitter, and LinkedIn, where we share breakdowns, the latest news, awesome artworks, and more.