Google Says Gemini 1.5 Pro Can Make AI Robots Smarter

Google’s Deepmind is training its robots with Gemini 1.5 Pro, and the results have impressed the researchers beyond any existing AI robot technology.

Google recently released demo videos showcasing the smart capabilities of its AI robots. Google’s dedicated AI research wing, Deepmind, is using Gemini 1.5 Pro to enhance the on-ground smartness of its RT-2 AI robots. The tech giant has been testing the robots for complex task handling against simple text prompt instructions. The Deepminds robotic team has released a research paper on the findings and multiple developments that can revolutionize the progress of AI robots.

The team filmed a video tour of a designated area and shared it with the robot to help it self-learn about its surroundings. A key element that helped the RT-2 AI robot was its longer processing capability using natural language instructions. The robot had a 90% success rate across 50 interactions in a 9000+ square foot operating area. 

It was able to perform complex tasks with self-added layers of smartness. For instance, a team member asked the AI robot if it could take him somewhere he could draw. Based on the video tour the robot learned on, it not only identified but also guided the person to the exact spot. The ability of the robot to understand extremely simple and conversational instructions is a breakthrough for on-ground implementation of AI technology.

Researchers also found preliminary evidence that Gemini 1.5 Pro can fulfill conversational instructions that go beyond just navigation. One employee who had multiple empty Coke cans on his desk asked the robot to find out if his favorite drink was available. The AI robot identified the drink through real-time observational learning and went to the refrigerator to check the inventory before answering the question. This added layer of self-awareness to the nuances of instructions is something the Deepmind team is investigating further on priority.

However, though impressive, the robots are still quite slow in processing the information. While the demo videos left out certain details, the research paper revealed that the RT-2 robots took a buffer time of 10-30 seconds before answering or taking action to the instructions. AI apps are achieving advanced milestones, like cracking the UPSC exam in just 7 minutes. With this breakthrough at Google, the possibility of bringing helper AI robots for household and home support will get on the fast lane.