New Agile ONE Humanoid Robot 21 DoF Tech Nano Banana Pro (AI NEWS)

This is the older version of the robot’s hand, and now looking at the newer version, it features this modular design, which means that each finger uses the same actuators, electronics, multi-sensory, and mechanical transmission systems, meaning that they can be interchanged with one another. So this is important because it not only simplifies the manufacturing process and maintenance, but it also allows for easier reconfiguration. And each one of these fingers features 10 newtons of active fingertip force, and a weight of about 1.5 kilograms per hand, with 360 degrees per second of joint velocity, and a 1 kilohertz control rate, as well as a fast-changing adapter flange for the robot’s hand.

And when it comes to the robot’s physical specs, it stands at 174 centimeters tall, or about 5 feet, 9 inches, and weighs 69 kilograms, or about 152 pounds, with a 20 kilogram carrying capacity, or about 44 pounds. Here you can see it’s turned on on the shoulder, and it features four cameras up in the head, which we’ll get into in just a second, as well as these eyes and an information display on its chest to give you real-time feedback. Here, the engineer speaks to it in English, using natural language to tell it to help him build an F1 car.

And you can see that it starts to walk to where it needs to do the work using its LIDAR and spatial recognition systems here. And this is all real-time feedback using its Agile 1 AI system for autonomous decision-making. Here, you can see it’s using its five-fingered hands to actually do these workflows, for which it was trained on proprietary data sets by Agile robots. And it’s already doing this real-time work. Check out these LIDAR cameras on the chest as well. What do you think those are? Two RGB-D cameras, or is it all LIDAR? And something else to mention about the fingers is that it has feedback sensors in its fingertips and palms to be able to feel exactly how much force to give to each one of these different processes.

And it delivers 10 Newtons of fingertip force for a very delicate but firm grip on objects in its surroundings, and 360 degrees per second of movement velocity to adapt in real-time to any situation. But physical AI isn’t the only space with breakthroughs, as Google just launched its Nano Banana Pro powered by Gemini 3 Pro, and it builds upon the predecessor Nano Banana, which was huge just a few months ago. And what sets this version apart is that unlike traditional image generators, Nano Banana Pro now integrates Gemini 3’s reasoning capabilities and real-world knowledge.

So the system can create these context-rich visuals like educational infographics, or transform handwritten notes into professional diagrams, and even access Google’s search knowledge base for real-time information like weather data or sports statistics. But it also features these new capabilities, including text rendering across multiple languages with unprecedented accuracy, and from short taglines to lengthy paragraphs with varied fonts and calligraphy, it features multilingual support for better content localization, and in terms of advanced consistency, it now can blend up to 14 images together while maintaining resemblance of up to 5 individuals, which can be useful for things like design mockups and brand consistency, and it also now offers these professional grade controls for localized editing, camera angle adjustments, sophisticated color grading, and lighting transformations, and the output supports various aspect ratios, plus 2K and 4K resolutions.

And in terms of availability, it’s already rolling out globally on Gemini app’s Create Images feature with the Thinking model, and free users can receive a limited quota before reverting to the original Nano Banana, while Google AI Plus Pro and Ultra Subscribers get higher access levels. Then, Google also just launched Gemini 3, which it claims is the world’s best model for multimodal understanding and its most powerful agentic coding system to date, which Google says marks a new era of intelligence on their path towards AGI, and meanwhile, to enhance its reasoning abilities, Google also introduced Gemini 3 DeepThink, which is an enhanced reasoning mode that pushes performance even further, and in testing, DeepThink surpassed Gemini 3 Pro’s already impressive benchmarks, scoring 41% on humanity’s last exam and 93.8% on GPQA Diamond, and most notably, it achieves an unprecedented 45.1% on Arc AGI 2 with code execution, which demonstrates extremely advanced problem-solving abilities for new challenges.

And then there’s Gemini’s multimodal integration across all content types. So it’s built to synthesize information across text, image, video, audio, and code, and Gemini 3 leverages a 1 million token context window, and it also now powers the Google Search AI mode, which generates these visual layouts, interactive tools, and simulations on the fly just based on user queries. And when it comes to coding and development breakthroughs, Gemini 3 also tops the WebDev Arena leaderboard at 1,487 ELO and scores 54.2% on terminal bench 2.0, which tests the model’s ability to operate computers via the terminal, and on the software engineering bench verified, it achieved a 76.2%, which measures the coding agent’s performance, and it’s a substantial improvement over Gemini 2.5 Pro.

But Google is also launching a new tool for agent-first development, and it’s called anti-gravity, which is this new agentic development platform that transforms how developers can work. And rather than serving just as another tool, anti-gravity agents now operate as active partners with direct access to editors, terminals, and browsers. And these agents can autonomously plan and execute complex and software tasks while validating their own code. So anti-gravity integrates into Gemini 3 Pro and Gemini 2.5 computer use model for browser control and nano-banana for image editing. But most impressive could be its long horizon planning capabilities, because Gemini 3 demonstrates improved long-term planning by topping vending bench 2, which basically tests models on managing a simulated vending machine business.

And Gemini 3 Pro even maintains consistent decision-making throughout a full simulated year of operation without task drift. And this translates to real-world applications where Gemini 3 can navigate complex multi-step workflows. And as for availability, Gemini 3 has already launched in preview across Google’s product suite, and Google AI Ultra subscribers can access agenda capabilities through Gemini Agent in the Gemini app with expansions to more Google products planned soon. And DeepThink is currently available to safety testers before being rolled out to Google AI Ultra subscribers. And developers can access Gemini 3 through Google AI Studio, Vertex AI, Gemini CLI.

Anyways, like and subscribe for more AI news and check out this video here. [tr:trw].