NVIDIAs 40000 Hour AI Experiment Changing Robotics Forever (AI NEWS)

So, in terms of the speed, it can walk at up to 6 kilometers per hour, or 3.5 miles per hour, which isn’t super fast, but you wouldn’t really want a robot being able to sprint after you, unless if it was for the use case that I’m about to show you with the next robot. But first, this one adopts a dual kernel mechanism, with 4 kilohertz communication frequency, and real-time core, as well as non real-time task isolation, which basically keeps everything smooth between computation and motor control. And the robot features a 1.01 kilowatt hour battery capacity for more than 2 hours of battery life.

Plus, the battery can be swapped out to avoid waiting for another charge. And in terms of power, the robot features maximum joint torque of up to 83 Newton meters. But the i2 is more like a starter robot, whereas their larger, newer humanoid, the i3, is what the company calls an intelligent mechanical guard. But this stands much taller at 180 centimeters, or about 5 feet 11 inches. So that’s about roughly a human male size. And as seen in the walking demo here, it does feature a biomimetic compliant gate, which just means that it walks just like a human.

And this is for multi terrain mobility, and running speeds of up to nine kilometers per hour, or over 5.5 miles per hour, which for reference, most humans jog between six to nine miles per hour. So could this really keep up with, say, someone running away from the cops? Or would it need to run significantly faster to catch up to them? And in terms of disturbance resistance, meaning that you can’t just push it over, it does feature a quite robust system to prevent it from toppling over, which also helps it with slope obstacle avoidance.

So the robot here is fitted with, as you can see, some type of a different helmet, it seems, and it’s holding a flashlight. So in terms of actual real world performance, the robot can serve in the real world in factories, as well as in laboratories and retail settings. And it’s able to tackle these real world tasks throughout these settings, thanks to its sub millimeter level precision accuracy when using real world tools. And it accomplishes this with a visual servo approach, which it uses to detect and align itself with screw holes for more precise fastening.

And in terms of position errors, it’s controlled within two millimeters using its AI vision and adaptive algorithms. And in garment factories, this robot has already successfully operated pattern sewing machines end to end using its adaptive grasping. And for electronics, it’s able to apply conformal coding to PCBs using its AI vision. And here’s the wild part. In terms of setup, it only requires CAD models and worker demonstration videos. So no production line modifications are needed. Basically, you just show it a video and a schematic, and it’s able to learn. And out of the box, it’s supposed to be compatible with existing hand tools across different coding targets.

So the other things that it can do is process data, do inspections, open doors, press buttons and test circuits. And it also supports natural language interaction for non technical personnel, meaning that you don’t need to be an engineer to tell it what to do, you’ll be able to just talk to it and it should be able to understand otherwise watch a video demonstration. And it features customizable voice gestures and facial expressions. But building this kind of robot hardware is only half the battle. Because these robots still need to learn about the real world.

And Nvidia just made a breakthrough by starting in the virtual one. So Nvidia just introduced NitroGen, which is its open source AI model that can play over 1000 games. And it learned by watching YouTube and Twitch streams. It was trained on 40,000 hours of gameplay footage with the researchers using controller inputs from videos where streamers displayed their buttons. And then they paired that with a computer vision technique and decoded exactly what the players were doing in order to create a massive data set of visual observations paired with actions. But what makes NitroGen really remarkable is its foundation because it’s built on Nvidia’s Groot N1.5 robotics architecture.

In fact, this is the first demonstration where models designed for physical robots can serve as universal agents across virtual worlds with completely different physics and visual styles. And when tested on unfamiliar games, NitroGen achieved 52% better success rates than models trained from scratch, which shows genuine cross game generalization rather than narrow title specific learning. But what if AI could also generate entire virtual worlds from scratch? Well, that’s exactly what researchers have achieved with Long V2. So Long V2 can generate ultra long consistent video environments lasting up to five minutes, which is a huge leap for video world modeling.

Because the challenge with AI video generation has always been that there’s a trade off, you can have control, visual quality or temporal consistency, but rarely all three. But Long V2 tackles this through a clever three stage training approach. So first, the model learns controllability by integrating both dense and sparse control signals, which gives users precise influence over what gets generated. Second, it addresses a common problem where AI video quality degrades over longer sequences through degradation aware training. And this bridges the gap between short training clips and extended inference. And this brings us to the third stage, which is history context guidance, and it assures that each generated clip aligns seamlessly with what came before it.

And this is what maintains coherent visuals across the entire video. And alongside it, researchers also released Long V Gen Bench, which is a new benchmark featuring 100 high resolution one minute videos, spanning real world and synthetic environments to properly evaluate these long form generations. And in testing, Long V2 already achieved state of the art performance across controllability, temporal coherence, and visual fidelity, all while generating continuous video of up to five minutes long. But what about real world agents? This is where Kyber comes into play as they just unveiled a highly dexterous robotic hand that can assemble parts and tighten screws at superhuman speeds.

So previously, workers thought that if they had a highly dexterous job, they would probably be safe from automation. But Kyber may be proving this wrong because they’re showing that doing things like placing parts, inserting screws and fastening them thousands of times daily could better be automated by these robotic hands because of the speed that they work at. And these are mounted on a collaborative robot arm, and the hand integrates seamlessly into any existing production line. But the real showcase is watching this robot work. This is in real time with no video edits.

And this is only possible because the hand is fully back drivable and torque transparent, meaning that it naturally adapts to the resistance of the nut rather than fighting against it. And for robotics developers, this matters beyond the visual impressiveness because back drivability enables simpler, more reliable manipulation for both traditional software control, as well as AI learning systems, which could potentially accelerate how quickly robots would master new tasks. But like and subscribe, check out this video here and thanks for watching. Bye. [tr:trw].