Google’s new robot AI can fold delicate origami, close zipper bags without damage

On Wednesday, Google Deep Mind Announced Two new AI models designed to control the robot: Gemini robotics and Gemini robotics-AR. The company claims that these models will help many forms and size robots understand the physical world and understand them in a more efficient and critical way than previous systems, which will pave the way for applications such as Humanoid Robot Assistants.

It is worth noting that although the hardware for the robot platform is moving at a steady pace (okay, Probably not always), Creating an AII model that can pilot these robots independently through safety and precision novel scenes. Which industry called “status AI” is a Moon shot round For example, Nvidia, and it is a sacred grill that can potentially transform robotics into common use workers in the physical world.

With these posts, new models of Google make it Gemini 2.0 Large language model Foundation, especially for robotic applications to include abilities. Gemini robotics has included Google in the “Vision Language Action” (VLA) capabilities, allowing visual information to act, understand language orders and create physical movement. On the contrary, Gemini Robotics-AR has focused on “statue reasoning” that enhances local understanding, and allows robotics to connect it to its existing robot control system.

For example, with Gemini robotics, you can ask a robot to “lift bananas into the basket”, and it will use the scene camera to identify bananas, and guide the robotic arm to successfully perform the process. Or you can say, “Fold an Oragami Fox”, and it will use your knowledge of Oriyagami and use the paper carefully to perform this task.

In 2023, we covered Google’s RT-2Which represented a remarkable move towards more common robotic capabilities using Internet data to help the robot understand language orders and adopt new scenarios, then doubled the performance on the tasks compared to its predecessor. Two years later, it seems that Gemini robotics have just made another coffee jump, not only to understand what to do, but also to implement complex physical manipulations that RT-2 cannot handle clearly.

Although RT-2 was limited to reproducing the already practicing physical movements, the Gemini robotics allegedly showed significantly enhanced skills that enable the first impossible tasks such as Oregi folding and zipper folk bags to pack breakfast. This change by the robot that only understands the orders of the robot that can carry out critical physical tasks shows that Deep Mind has begun to solve one of the biggest challenges of robotics: transforming the robot into a cautious, precise movement in the real world.

Better normal results

According to Deep Mind, the new Gemini robotics system shows a very strong general, or shows the ability to perform novel function that was not trained to do specifically than its previous AI models. In its announcement, the company claims that Gemini Robotics “is more than doubles performance on a comprehensive general benchmark than the latest vision language action model.” It is important to make it common because robots that can adopt new scenario without specific training for each situation can one day work in an unexpected real -world environment.

This is important because doubts remain on how useful the Humanoid robots can be or how much they are really capable. Tesla Exposed Last October, its Optims General 3 robots claim the ability to complete many physical tasks, yet concerns remain on the authenticity of its independent AI capabilities after the company. Enrolled That its spraying demo was controlled by many robots by humans.

Here, Google is trying to make the real thing: a common robot brain. Keeping this goal in mind, the company announced a partnership with Texas -based Austin Epitronic “To build the next generation of Humanoid robots with Gemini 2.0.” While mainly trained on a Baiminuel Robot platform called Woha 2Google says Gemini can control various robot types based on robotics research Franka Robotic Arms For more complex hypnrops such as epitronic Apollo robots.

Although the Humanoid robot approach is a relatively new application for Google Generative AI models (from this cycle of LLM -based technology), it is worth noting that Google had previously acquired several robotics companies around 2013-2014 (including. Boston dynamicsWhich makes humanoid robots), but later Sold them. New partnerships with epitrophone seem to be a fresh approach to humanoid robotics, rather than direct continuity of these previous efforts.

Other companies are working hard on the Humanoid robotics hardware, such as data AI (which Unharmed The main funding for its Humanoid robot in March 2024) and Boston Dynamics, a subsidiary of the aforementioned alphabetical alphabet (who Introduced Last April, a flexible New Atlas Robot), but a useful AI “driver” has not yet revealed the robot to make the robot really useful. On this front, Google has also had limited access to the Gemini Robotics-AR through the “reliable tester” program, such as Boston Dynamics, Aging Robotics, and Enchanting Tools.

Safety and limits

Google, Google, has mentioned a “layered, comprehensive approach” of safety protection, which maintains traditional robot protective measures such as collision and maintains the limits of strength. The company describes the development of A “Robot Constitution“Framework influenced by Isaac Asimov Three rules of robotics And issuing a dataset is surprisingly called “Asimov“Help researchers evaluate the safety implications of robotic actions.

This new Asimov Dataset represents an attempt to create standard methods to evaluate the protection of the robot from the prevention of Google’s physical damage. Datastate researchers are designed to help test how to better understand the potential consequences of robot measures in various landscapes. According to Google’s announcement, Dataset “will help researchers strictly measure the safety implications of robotic actions in real -world scenarios.”

The company did not announce availability timelines or specific trade applications for new AI models, which are pending in the research phase. Although Google has jointly presented progress in the AI-driven capabilities in the demo videos, the controlled research environment still leaves open questions about how this system will actually perform in the unexpected real-world settings.