AI from the giant NVIDIA is revolutionizing the way robots learn, allowing them to learn and perform complex tasks, almost like magic tricks. Man and robotics have never been so close.
Eureka, this is the name of the new artificial intelligence agent created by the NVIDIA Research teams. The AI uses the company’s Large Language Models (LLMs) to automatically generate reward algorithms, which help train robots to complete complex tasks, like spinning a pen quickly. Seeing AI teach skills to robots is as stunning as it is promising.
With its new AI agent, NVIDIA revolutionizes robotic learning
The new AI agent, developed by NVIDIA, has reached a stunning milestone in the field of robot learning. Eureka, powered directly by the powerful GPT-4 language model, taught a robotic hand to perform rapid pen turns, with human-like precision.
Imagine, dear readers, that this feat is only one of around 30 tasks that the robots have been able to master thanks to Eureka, which autonomously writes reward algorithms to train them.
Eureka revolutionizes reinforcement learning here, by teaching robots to perform complex tasks, for example opening drawers and cupboards, throwing and catching a ball, but also handling scissors. And all this, without requiring predefined reward models, a major breakthrough.
An improvement in robot performance of more than 50%
Anima Anandkumar, senior director of AI research at NVIDIA, points out that “ Reinforcement learning has achieved impressive victories over the past decade, but many challenges remain, such as reward design, which remains a process of trial and error “.
Eureka is able to solve these challenges by outperforming reward programs written by experts, in more than 80% of tasks. We then see an average improvement in performance of more than 50% for the robots.
What lies behind Eureka
Eureka exploits, as we said, the power of generative AI and OpenAI’s GPT-4 language model, in particular to write algorithms. Behind this, GPU-accelerated simulation in Isaac Gym (NVIDIA’s physics simulation environment, an end-to-end high-performance robotics simulation platform), allows Eureka to quickly evaluate a large number of reward candidates, which helps optimize training effectiveness.
Once the training results are generated, the agent asks GPT-4 to improve its generation of reward functions, contributing to the continuous improvement of the AI itself! Eureka is already capable of teaching a whole variety of robots: bipeds, skilled robotic hands, quadrupeds or even cobot arms.
Eureka’s research paper features detailed assessments of 20 AI-trained tasks, each time demonstrating the robots’ ability to perform a diverse range of complex manipulation skills. By combining NVIDIA’s large language models and GPU-accelerated simulation technologies, Eureka opens the door to creating new realistic physics animations for artists.
In the genre, NVIDIA already operates Voyager, an AI agent capable of playing autonomously Minecraft. And the potential is immense: computer vision, self-driving cars, computer graphics and robotics, major technology players are redefining the way robots learn and interact with the world around them. Eureka is the latest achievement of this innovative research.
The emergence of artificial intelligence as a mainstream tool has opened up numerous possibilities for all content producers. Text, image, sound… This new fashionable technology can now provide assistance in many areas, and facilitate work in the most difficult stages of creation. And with an ever-increasing offering, it is important to distinguish which tools provide real added value. So you don’t waste hours trying everything the Google results pages offer!
Read more
Source : Eureka-Research (NVIDIA)
2