Skild AI Learned 100000 Robots In 1000000 Simulation Years To Do THIS

And the N1 has 23 movable joints with custom actuators to allow it to run at up to 7.8 miles per hour or about 12.5 kilometers per hour. Plus it can climb 20 degree slopes, ascend 8 inch stairs, and balance on one leg. As for its power, it’s a plug-in battery that powers it for over two hours continuously. And Fourier says that N1 has been tested for over 1,000 hours outdoors and 72 hours non-stop. And soon, Fourier says they’ll release AI inference code and training frameworks for vision, speech, and multimodal integrations.

But the question is, how much should it cost and should it be for purchase or just for lease so the company still owns the data? But robot hardware is only half the equation because the real breakthrough is teaching robots how to see and adapt like humans do. Which is why Stanford researchers just announced Visual Mimic, their framework which now allows humanoid robots to perform complex movements and manipulation tasks using only visual input, meaning no motion capture or external tracking systems required. And to start, Visual Mimic system uses a hierarchical policy architecture with two components.

The first is a low-level tracking policy, the second is a high-level generation policy. And both of these AI policies are trained entirely in simulation using reinforcement learning and behavioral cloning. And then they’re deployed on physical robots without any additional real-world training. And in the real world, Visual Mimic demonstrated zero-shot generalization across multiple conditions, including different locations across Stanford’s campus, varying lighting conditions throughout the day and night, and tasks requiring whole body coordination, including box kicking, football manipulation, and full body lifting. And because Visual Mimic relies exclusively on egocentric vision using cameras mounted on the robot, it also solves a central challenge in robot research, which is using reinforcement learning to produce natural human-like behavior.

In fact, the team embedded human motion priors into the robot’s low-level policy through distillation, plus they regularized the action space using human motion statistics in order to prevent unstable training. And there was an unexpected finding through all this. When trained with varying ground friction, the AI policy actually learned to adaptively use different body parts for contact while maintaining human-like movement patterns. But what if the robot has jammed motors, broken limbs, or encounters unexpected conditions? How can it adapt on the fly? Well, skilled AI researchers just solved this problem by training a single AI model to control 100,000 different simulated robot bodies over 1,000 years of simulated time, and the result is an omnibody brain that adapts to scenarios that it never trained on.

This is called zero-shot adaptation, and it all happens without any fine-tuning, as demonstrated across the following scenarios. The first of these is learning from failure, where a quadruped with passive leg knobs fell twice before successfully adapting its strategy on the third attempt, and this shows in-context learning that’s similar to large language models. The second situation was that of missing limbs. When a robot’s calf was cut, removing four degrees of freedom, it struggled initially but discovered within 7-8 seconds that swinging its thighs allowed it to regain control. The third situation is broken joints, with knees creating a three-legged robot, but skilled AI brain shifted its weight backwards and walked after two to three seconds of adaptation.

The fourth situation was jammed wheels. When the wheels were locked mid-motion, the AI instantly switched from rolling to walking, and then it went back onto its wheels once they were unlocked again. And the fifth situation was stilts, where longer legs raised the center of mass beyond the original training data, but the AI quickly adjusted its step timing and placement to stabilize its walking. But artificial intelligence is also adapting in the 3D world with the release of a new model called P3-SAM, and it can automatically break down any 3D object into its constituent parts, which addresses long-standing limitations in 3D asset segmentation that have hindered applications ranging from model reuse to generative design.

And here’s the problem. Segmenting 3D assets into meaningful parts is fundamental for 3D understanding, and this enables various downstream applications like part level generation and manipulation. But until now, AI has struggled with complex objects, many times requiring significant manual intervention, which limits their practical utility and automated workflows. But P3-SAM works differently by using META’s segment anything model for 2D images, plus P3-SAM brings point promptable segmentation to native 3D data with its architecture consisting of three main components. The first of these analyzes the 3D object shape and structure, translating its geometry into data that the AI can understand, like creating a detailed map of every point on the object’s surface.

Then the second component generates multiple predictions after different detail levels. Think of zooming in and zooming out on an object, and this allows the system to identify both large components like car doors or small parts like the door handle in the same bath. Next, the third component checks which prediction is most accurate by comparing how well each segmentation matches the actual object boundaries, and then it selects the best one. And the system works automatically by selecting strategic points across the 3D object, and then it runs them through P3-SAM to identify potential part boundaries.

And the model generates multiple overlapping predictions, then filters out duplicates to keep only the most accurate segments. Finally, these segments are mapped onto the object’s surface to define where each part begins and ends, completing the breakdown without human input. And for custom segmentation, users can even interact with the system by providing their own point prompts. Meanwhile, Google Cloud is unifying AI models as FAL just released its platform of over 200 AI models using a single API to streamline multimodal integration. And the platform central hub aggregates both open source and proprietary models so that developers don’t need to manage multiple API integrations.

On top of this, developers can chain together multiple models to combine video, image, and audio processing and build complex media production pipelines without custom infrastructure. And when new models are released, FAL’s API structure makes upgrades and switching between models easy. But Alibaba is pushing the envelope even further with the release of Quen3VL, its multimodal AI that significantly expands visual understanding and agent capabilities, and it comes in two architectures, dense and mixture of experts, with instruct and thinking additions for different use cases. And its key capabilities include first its visual agents, meaning Quen3VL can autonomously operate computer and mobile interfaces, recognizing UI elements, invoking tools and completing tasks.

And it also generates functional code in HTML, CSS, JavaScript and more directly from images and videos. Its second key capability is its extended content window of up to 1 million tokens for processing full books and hours long video with precision. And its enhanced 3D spatial reasoning can even judge object positions and occlusions for robotics applications. Its third key capability is expanded pre-training to recognize celebrities, products, landmarks and more. Plus its OCR now supports 32 languages with better performance in low light blur and complex document structures. And its final capability is STEM reasoning, having upgraded mathematical and logical reasoning for casual analysis and evidence based answers that compete with specialized models.

But like and subscribe for more AI news and listen to Microsoft’s newest text to audio AI called Vibe Voice that can generate up to 90 minutes of conversation with four speakers at 80 times the efficiency. I can’t believe you did it again. I waited for two hours. Two hours! Not a single call, not a text. Do you have any idea how embarrassing that was just sitting there alone? Look, I know I’m sorry. All right. Work was a complete nightmare. My boss dropped a critical deadline on me at the last minute.

I didn’t even have a second to breathe, let alone check my phone. A nightmare? That’s the same excuse you used last time. I’m starting to think you just don’t care. It’s easier to say work was crazy than to just admit that I’m not a priority for you anymore. [tr:trw].