NVIDIA Research has made a new AI tool called Eureka that teaches robots to do complex tasks. It taught a robot how to do quick pen spin like a human for the first time.
According to the official blog post of NVIDIA, it is just one of the 30 tasks the robots have learned to perform expertly. Besides pen spinning, it has taught many tasks to robots, such as opening drawers, manipulating scissors, and tossing and catching balls.
Eureka teaches robots by writing reward algorithms to train bots autonomously. The tool leverages the cutting-edge capabilities of GPT-4’s large language model (LLM) and NVIDIA’s own Isaac Gym, a development platform for building 3D tools and applications.
Importantly, this tool works automatically and does not need human prompting or pre-defined templates. It teaches robots by formulating a reward algorithm specifically tailored to the task and morphology of the robot. The reward scheme allows robots to learn by trial and error.
When tested against human specialists in 29 different environments using ten different robot platforms, Eureka produced superior reward programs. The official paper of NVIDIA mentions this tool leads to an average performance improvement of more than 50% for the bots.
The code and benchmarks of Eureka will be open-source. This will allow developers to use resources to expand on the research and leverage the community. Recently, it released a video that shows the workings of this tool.
The introduction of Eureka by NVIDIA has the potential to revolutionize robot learning by providing a platform for robots to learn complex skills with high efficiency. As the technology develops more, we expect an increasing integration of these AI-driven robots into various industries.