Robot hands can match human dexterity with new AI smarts, claims Navia

Navia researchers have achieved great success in robotics EurekaAn AI agent that is said to teach bots complex skills like a pen twirling human.

The new technique, described in a paper published Thursday, builds on recent advances in large-scale language models such as OpenAI's GPT-4. Eureka uses generative AI to autonomously write sophisticated reward algorithms that allow robots to learn through trial and error reinforcement. The paper states that this approach is 50% more effective than human-written programs.

“Eureka has taught quadrupeds, dexterous hands, cobot arms and other robots to open drawers, use scissors, catch balls and about 30 different tasks.

Eureka is the latest demonstration of Nvidia's pioneering work in driving AI through language models. Recently, the company open-source SteerLM-Mechanism to train AI assistants on human feedback to make them more helpful.

Similar to Eureka, SteerLM uses advances in language models, but focuses on a different challenge—improving AI assistant alignment. SteerLM trains assistants to practice conversations like a learning robot. The system provides feedback on assistant responses with attributes such as helpfulness, humor, and quality.

For example, it's like a robot learning to dance from videos labeled good or bad, rather than having a human eye review thousands of random dances and choosing which ones are good or not (which is how your typical AI chatbots are trained). Through repeated practice and feedback, the assistants learn to provide responses tailored to the user's needs. This will help make AI more useful for real-world applications.

The common thread is using advanced neural networks in creative new ways to teach robots or chatbots. Nvidia is pushing the boundaries on both the hardware and software fronts.

The key to Eureka was combining simulation technologies. Isaac Jim With the pattern recognition capability of language models. Eureka effectively “learns to learn,” optimizing its own reward algorithms from multiple training runs. It even accepts human input to verify the reward.

This self-improving approach has been proven so far in training all kinds of robots – legged, wheeled, flying and dexterous hands.

Nvidia's Eureka and SteerLM aren't just breaking barriers, they're teaching robots and AI fine art and intelligent interaction. With every pen and comic conversation, you're painting a future where AI doesn't just look like it, but creates a new one with us.