NVIDIA Eureka is an AI agent developed by NVIDIA Research that autonomously generates reward algorithms to train robots in performing complex tasks. Leveraging OpenAI's GPT-4 large language model, Eureka creates software code enabling robots to learn through trial and error, significantly enhancing their performance.

In evaluations across 29 tasks, Eureka's reward functions outperformed expert human-written ones in 83% of cases, with an average performance improvement of over 50%. Notably, Eureka has trained a robotic hand to execute rapid pen-spinning tricks with human-like proficiency and has enabled a quadruped robot to balance and walk atop a yoga ball. This advancement underscores the potential of AI-driven reward design in elevating robotic dexterity and learning efficiency.