Nvidia launches AI system Eureka, which allows robots to perform complex tasks such as pen-turning and passing

News on October 23, NVIDIA recently announced an AI system called Eureka, which is based on OpenAI’s GPT-4 and allows robots to perform tasks such as “turning a pen”, “opening a drawer”, and “holding scissors” ", "passing the ball between hands" and more than 30 complex actions.

After inquiry, this site learned that NVIDIA Research Institute led the development of Eureka. This AI system allows developers to cooperate with NVIDIA's own physics simulation software Isaac Gym for reinforcement learning (reference learning).

Anima Anandkumar, senior director of NVIDIA's AI research department, believes that although "reinforcement learning" has made progress in the past 10 years, there are still many challenges, such as "reward design" and other aspects, which are still stuck at " trial and error” phase. The Eureka currently announced by NVIDIA is designed to perform difficult tasks and is the first attempt to combine generative AI and reinforcement learning algorithms.

The "reward design" scheme generated by Eureka using GPT-4 can assist the robot's "trial-and-error" learning, and can take over the tasks of 80% of human experts, thus enabling The average training efficiency of robots has increased by more than 50%.

It is reported that when Eureka generates a "reward design" plan, developers do not need to enter additional task prompts or write predefined reward templates. Combined with manual correction of rewards, the robot's actions are ultimately more in line with the developer's intentions. .

When the AI ​​system is used with Isaac Gym, it can use GPU acceleration to conduct simulations and quickly evaluate the quality of a large number of batches of candidate feedback, thereby improving training efficiency and making statistical summaries based on training results to improve " "Reward Scheme", while improving the robot's response level, makes the AI ​​system applicable to a wider range of robot types, such as "animal type" (four-legged), "humanoid (two-legged)", "four-legged" Axis aircraft, arm-type robots and other robots can be equipped with the Eureka system.

▲ Picture source NVIDIA

▲ Picture source NVIDIA

NVIDIA will " Open Source Sensitivity Benchmark Test", after evaluating the training results of 20 tasks, obtained satisfactory results.

Accordingly, NVIDIA believes that Eureka can achieve dexterous robot control and provide animators with new ways to "generate realistic physical movements."

The above is the detailed content of Nvidia launches AI system Eureka, which allows robots to perform complex tasks such as pen-turning and passing. For more information, please follow other related articles on the PHP Chinese website!