Google Just Unveiled Gemini Robotics and It's Whole New Level of AI Robot Intelligence!

Google DeepMind just introduced two powerful AI models called Gemini Robotics and Gemini Robotics ER. And they're designed to take robots far beyond anything we've seen before. These models can see the world, understand your instructions and take action. It's actually giving robots a whole new set of instincts and it's already showing results that could change how we think about automation. Alright, so on March 12, 2025, Carolina Parada at Google DeepMind announced the launch of Gemini Robotics, which is an advanced vision language action VLA model built on top of Gemini 2.0. The reason this is interesting is that Gemini 2.0 was already recognized for being super powerful at multimodal reasoning. Things like reading text, interpreting images, and even parsing audio and video. But for the most part, that was just about digital data. Right? It didn't really address how a robot might fit, physically move and manipulate objects in real life. Now with Gemini Robotics, we get a system that can directly control robots by generating physical actions as an output, not just text or code. But that's not all. Google DeepMind has also introduced a sister model called Gemini Robotics ER. ER in this case stands for embodied reasoning, which means it has advanced system spatial understanding, basically the ability to figure out how objects exist in physical space, how they might move, how they can be grasped, or how to plan a path for a robotic arm to grab something. It's like giving the AI an extra layer of problem solving superpowers that are really specific to real world tasks. So let's get into some of the specifics. Gemini Robotics is described as a vision language action model. That means it sees what's going on around it like a video feed from a camera. Then it takes in instructions in natural language so you can just talk to it almost like you'd talk to a person. And finally, it produces an action plan that controls a robot. For example, imagine you want the robot to fold a piece of paper into a fancy origami shape, or maybe pack a snack into a Ziploc bag. Gemini Robotics can handle these tasks even if it hasn't been explained explicitly trained on that exact scenario. And that's a big shift from traditional industrial robots, which require a lot of manual programming to do something new. Google tested Gemini Robotics on a generalization benchmark to see how well it handles tasks that it wasn't explicitly taught. Turns out it more than doubled the performance of previous state of the art models in this category, which is basically a huge leap. In generality, that means it's better at dealing with new objects it hasn't seen before. New types of instructions and even changes in the environment. Like if something gets moved or slips out of its grip. In a dynamic real world setting, that's super important because, let's face it, things rarely go according to plan. Another area where Gemini Robotics shines is interactivity. Robots operating in a chaotic environment have to adapt quickly. If a person, say, picks up an object or moves something out of place, Gemini Robotics continuously monitors everything around it and replans on the fly if the environment changes. Plus, since it's built on top of Gemini 2.0, it's got these advanced language understanding capabilities, so it can process casual commands in everyday language. And if you say something like, could you move those boxes from the table to the shelf? And by the way, keep the red box on the floor, the the model can figure that out. That's pretty sweet for anyone who wants a robot assistant in the home or in a workplace setting. Now, Gemini Robotics is also about dexterity. One of the hardest challenges for robotics is doing the things we humans take for granted, like carefully manipulating objects. Folding origami, Packing a tight bag, or even turning a doorknob can be surprisingly tricky for a robot. Gemini Robotics, though, has been shown to handle really fine motor tasks. Google DeepMind even has these short clips of the robot's manipulator arms, folding paper or picking up pretty fragile items without wrecking them. That level of precision is a big part of why people are calling this model a game changer. Then we have Gemini Robotics er. Basically, if Gemini Robotics is the direct agent that controls the arms and does the tasks, Gemini Robotics ER is more like the brains behind advanced sports spatial reasoning. It looks at objects, figures out where they are in 3D, identifies the best way to grasp them, and even plans a safe trajectory for the robot to take. Imagine, for instance, you have a coffee mug on the table and you want the robot to pick it up by the handle from just the right angle. That's exactly the sort of nuanced scenario Gemini Robotics ER can handle. Here's the kicker. Gemini Robotics ER can do all of this right out of the box. Perception state, state estimation, spatial understanding, planning, code generation, without needing a bunch of separate modules, that alone is kind of a big deal because it streamlines the process for roboticists. In fact, in an end to end setting where the model is taking care of everything from vision to motion planning, Gemini Robotics ER scores two to three times higher in success rate compared to the baseline Gemini 2.0. And yes, it can generate robot configuration scripts using Gemini 2.0's coding abilities. If a task is still too hard, you can just show the model a handful of examples of how to do it, and it learns from those demonstrations. Now let's talk a bit about how Google is planning to get these models out into the real world. They've partnered with Apptronic, a company that specializes in humanoid robots, particularly the Apollo robot platform platform. And Google's not just collaborating with Apptronic, they're also an investor. Apptronic apparently raised $350 million last month with Google as part of that. That's a good sign that we might see humanoid robots powered by Gemini 2.0 in the near future. But it's not just Apptronic. Other big names in robotics like Agile Robots, Agility Robots, Boston Dynamics and Enchanted Tools are also involved as trusted testers of Gemini Robotics. Er, so these models are being tested in multiple environments and robot types, which should accelerate real world adoption. Interestingly, they're also introducing a new dataset to help measure and improve semantic safety in embodied AI and robotics. They're calling it the Asimov Data Set, in reference to Isaac Asimov's famous Three Laws of robotics. Part of the reason for the name is that Google has been experimenting with a robot constitution, basically a set of natural language rules that the model follows to steer it away from unsafe or unethical tasks. They say they've developed a framework to automatically generate these constitutions so the robot can be given specific instructions that align with human values. The Asimov Data set will help researchers test how well these rules actually work in real world robot scenarios and see how the model deals with moral or safety based decisions. Another important layer is Google's internal oversight. They have a Responsibility and safety council that reviews how AI is being developed. They also talk with outside experts about the societal implications of advanced embodied AI. That's key because let's be real. Giving robots the power to act on their own raises importance important ethical and safety questions, especially as these systems become more general purpose and are placed in settings like homes or workplaces. Now, from the user's perspective, what's really cool is the ease of use. If you have Gemini Robotics running on a new robot, you might just talk to it in everyday language. Pick up that box, then go ahead and place it in the corner. You don't have to know any special programming language and the robot can handle exceptions in real time, like if the box is hidden heavier than expected or if someone shifts it a few inches, messing up the robot's initial plan. This is an entirely different approach from the old days of meticulously coding movement patterns line by line for roboticists who still want a high level of control, Gemini Robotics er can generate Python or other languages to configure the robot's low level controllers, and you can tweak that code if you want. If you show it examples of how you want something done done, it can pick up that style of programming or approach to movement and replicate it in new contexts. That's where the model's instexed learning abilities shine. If you're as hyped about robotics as I am, I'm sure you'll be keeping a close eye on further demos and real world tests of these systems. I have a strong feeling we'll see some impressive use cases popping up soon, especially with Apptronics, Apollo Robot and others from Boston Dynamics, Agile Robots and beyond. This might just be the next big step in making robots genuinely helpful, user friendly and safe in everyday situations. Anyway, let me know your thoughts and as always, thanks for watching and I'll catch you in the next video.