Scientists from NVIDIA, originating from the University of Pennsylvania, have developed a new AI agent designed to train robots. The algorithm, named Eureka, utilizes the famous GPT-4 language model, which also powers ChatGPT. This algorithm would be used to enhance the performance of robots both in terms of reasoning and motor skills.
Eureka develops reward programs that help improve robots’ skills and enable them to perform complex tasks. This is essential for robots to be useful in everyday life, as they need to move their bodies in difficult and efficient ways in challenging environments.
To demonstrate this, NVIDIA has presented a video showcasing the nearly 30 skills that robots have developed thanks to Eureka. One of them is a robotic hand that performs tricks with a pen. To be honest, the agility of this hand is nothing to be envious of compared to a human. Other skills that Eureka has been able to teach robots include opening drawers, throwing and catching balls, and much more.
How does Eureka train the robots?
One of the processes used to train a robot involves having it perform tasks such as moving or grabbing objects within a physical environment. Every time the robot fails one of these tasks, it “learns” from its mistakes to avoid them in the future. However, NVIDIA researchers have tested a much more “digital” approach.
To do this, they have employed virtual environments similar to those in video games. Thanks to the Isaac Gym platform, it is possible to simulate real-world physics, allowing robots to train their abilities within virtual worlds. Isaac Gym is based on NVIDIA Omniverse, a real-time 3D graphics collaboration platform.
As stated in NVIDIA’s official blog, “the AI agent uses GPT-4 and generative AI to write code that rewards and reinforces learning for the robots.” Moreover, the AI can improve itself based on the test results.
