NVIDIA's New AI Model Enables Robots to Master Pen Spinning Tricks Like Humans

The role of humans in robotics, even as educators, is diminishing due to the rise of AI. NVIDIA Research has introduced Eureka, an AI agent powered by GPT-4 that has successfully trained robots to perform various tasks using reward algorithms. Remarkably, Eureka has taught a robotic hand to perform pen spinning tricks as well as, if not better than, many humans.

Eureka has also enabled quadruped robots, dexterous hands, and collaborative arms to accomplish nearly 30 tasks, including opening drawers, using scissors, and catching balls. According to NVIDIA Research, the agent's trial-and-error reward programs are 80% more effective than those designed by human experts, resulting in over a 50% improvement in robot performance. Additionally, Eureka self-evaluates its training outcomes, making adjustments to reward functions as necessary.

NVIDIA Research has made its Eureka algorithms available in a library, inviting others to experiment with them using NVIDIA Isaac Gym, the organization’s physics simulation reference for reinforcement learning.

The concept of robots teaching robots is gaining traction. A May 2023 paper in the Transactions on Machine Learning Research journal introduced SKILL (Shared Knowledge Lifelong Learning), a system that enabled AI systems to learn and master 102 distinct skills, such as diagnosing diseases from chest X-rays and identifying flower species. These AIs acted as teachers, sharing knowledge over a communication network. Additionally, researchers from institutions like MIT and the University of Bristol have made strides in using AI to instruct robots in object manipulation.