The NVIDIA Research team has introduced an innovative AI agent capable of instructing robots in intricate skills, exemplified by its ability to perform the challenging task of high-speed and precise pen spinning. The video below showcases how this robotic hand's performance closely aligns with that of a human.
_1697948471.jpg)
Eureka, an AI agent, can automatically create reward algorithms that teach robots new skills. It has helped robots master nearly 30 different tasks, including opening drawers and cabinets, throwing and catching balls, and using scissors.
According to the paper, Eureka's generated reward programs outperform human-written ones on over 80% of tasks, resulting in an average performance boost of more than 50% for the robots.
Eureka collaborates with the GPT-4 LLM and generative AI to produce software code that rewards robots for reinforcement learning. It doesn't require task-specific instructions or predefined reward templates but instead adapts to the developer's objectives and preferences. The AI agent also incorporates human feedback to fine-tune rewards, aligning them more accurately with the developer's vision.
Eureka leverages GPU-accelerated simulations in Isaac Gym to efficiently assess and optimize reward functions across various robotic tasks and domains. Afterwards, it compiles key statistics from the training results and instructs the LLM to enhance its generation of reward functions, making the AI self-improving. Eureka has successfully trained a wide range of robots, including quadrupeds, bipeds, quadrotors, dexterous hands, cobot arms, and others, to accomplish diverse tasks.
Photo: Reuters
© Copyright 2023. All Rights Reserved Powered by Vygr Media.
