Facilitating the Spread of Knowledge and Innovation in Professional Software Development

Write for InfoQ


Choose your language

InfoQ Homepage News Nvidia Introduces Eureka, an AI Agent Powered by GPT-4 That Can Train Robots

Nvidia Introduces Eureka, an AI Agent Powered by GPT-4 That Can Train Robots

Nvidia Research revealed that it has created a brand-new AI agent named Eureka that is driven by OpenAI's GPT-4 and is capable of teaching robots sophisticated abilities on its own.

With the help of this new AI, robots may be taught intricate feats like spinning pens in a manner akin to how people learn. Robots can learn through trial and error reinforcement thanks to Eureka's intelligent reward algorithms, which are created using generative AI and big language models like OpenAI's GPT-4.

This method has been shown to be almost 50% more effective than conventional human-authored programs, according to a paper written by Nvidia. According to a post on Nvidia's official blog, Eureka has been effective in teaching robots to perform a variety of activities, such as opening drawers, using scissors, catching balls, and more.

Anima Anandkumar, senior director of AI research at NVIDIA quoted:

Reinforcement learning has enabled impressive wins over the last decade, yet many challenges still exist, such as reward design, which remains a trial-and-error process.

According to the research, Eureka-generated incentive schemes outperform skilled human-written ones on more than 80% of tasks because they allow robots to learn through trial and error. The bots' performance improves by an average of more than 50% as a result.

Robots are rewarded for reinforcement learning by the AI agent, which uses the GPT-4 LLM and generative AI to create the necessary computer code. It doesn't require task-specific prompts or pre-defined incentive templates, and it readily accepts human feedback to change its rewards for outcomes that are more closely in line with a developer's goal.

The major innovation for Eureka was the fusion of language models' capacity for pattern detection with simulation technologies like Isaac Gym. Eureka successfully "learns to learn" by fine-tuning its own reward algorithms after a number of training cycles and even taking into account human input.

This research complements recent innovations from Nvidia Research, such as Voyager – an AI agent powered by GPT-4, capable of independently engaging in Minecraft gameplay.

About the Author

Rate this Article