The Nvidia research team has developed an AI system that can teach robots more complex tasks. The researchers called their new model Eureka.

Researchers at Nvidia Research have developed an AI system that teaches robots to perform complex tasks via LLMs. The project is called Eureka and writes its own reward algorithms that are used to train the robots. Eureka has now taught various robots to perform almost thirty more complex tasks. After the news of an EU investigation into anti-competitive practices, it is once again positive news for the chip manufacturer.

Helpful hands

Some of the tasks mentioned by Nvidia Research don’t seem that complicated to people:

• open drawers and cupboards
• throw and catch a ball
• Use scissors.

However, this is currently not so easy for a robot.

What the team is obviously proud of is a robotic hand that has been taught to perform tricks with a pen. Nvidia claims the robotic system can do this just as quickly as a real human.

Fitness for AI

Nvidia Research publications contain algorithms that developers can use. You can use Isaac Gym, a simulation environment that Nvidia created based on its Omniverse platform. This is used to create 3D tools and applications.

Eureka itself is powered by OpenAI’s GPT-4. According to Anima Anandkumar from Nvidia Research, the project is a first step towards new algorithms that integrate new learning methods to solve more complex tasks.

Machine learning

According to Nvidia Research, the algorithms written by Eureka performed 80 percent better than those written by humans. This would lead to the robots’ performance being halved.

Through generative AI and LLMs, Eureka does not require specific prompts or templates to write its software code. Human feedback is also directly integrated.

The GPU-accelerated simulation in Isaac Gym allows Eureka to quickly evaluate for more efficient workouts. The system then summarizes the statistics so that the LLM can improve its reward functions. This is how AI improves. As for GPUs, Nvidia also introduced Blackwell, the successor to Hopper, earlier this month.

Undoubtedly, Archimedes could never have predicted this when he released his iconic exclamation to the world from a bathtub.