The name of the base system (platform): | GPT-4 (neural network) |
Developers: | Nvidia |
Date of the premiere of the system: | October 2023 |
Branches: | Information Technology |
Technology: | Robotics, Application Development Tools |
2023: Product Announcement
On October 20, 2023, the research division of Nvidia Research announced the development of an artificial intelligence-based Eureka system designed to train robots in complex skills. As a result, machines are able to perform some actions even better than people.
Eureka is based on a large GPT-4 language model developed in OpenAI. Training is provided in the Nvidia Isaac Gym application, designed to simulate physical processes. At the same time, the neural network independently writes reward algorithms for training bots, which in this case learn skills by trial and error. Nvidia says Eureka's incentive programs outperform experts' peers in more than 80% of tasks. As a result, the performance of bots increases by an average of more than 50%.
Reinforcement training has made impressive progress over the past decade (by 2023), however some challenges still persist. Eureka is the first step towards developing new algorithms that combine generative artificial intelligence and reinforcement training to solve complex problems, says Anima Anandkumar, senior director of AI research at Nvidia. |
During the experiments, the Eureka system trained the robotic arm to perform quick tricks with the rotation of the finger handle. In addition, this AI agent made it possible to quickly train machines in other actions, such as manipulating scissors, throwing the ball, opening/closing boxes and cabinets, etc. Eureka receives a summary of the learning process and gives an indication of the language model for optimizing reward functions. So AI is self-perfecting.[1]