Unitree G1 Anti Gravity HDMI Humanoid New Kangaroo AI Robot Wan2.2 Animate SpatialGen 3D

Specifically, it’s able to bend 90 degrees forward onto its knees right there to put itself back on its feet. Here’s where it gets super impressive. See, the robot gets kicked over, it hops back up. Now it starts with its spinning moves. It starts jumping on its head. Here we can see it starts to do various types of kung fu moves or tai chi. Tell me in the comments exactly what it is. Looks like it’s getting ready to fight back. Here it’s doing spins. We’ve never seen any robot do these types of moves before, that’s for sure.

And we’re certainly approaching a level where these could not only start fighting each other, but maybe even face off with humans soon enough. So tell me in the comments whether or not you think that humans need certain regulations to prevent these robots from becoming too agile, too capable of fighting back potentially in the future. Look at these flips here. Quite impressive the way that it just seamlessly holds its balance and repeats. This is all using its latest reinforcement learning AI policies. Next, we’re going to talk about something that could completely transform how humanoid robots learn to interact with the world and us.

So we’ve all seen robots that can walk and maybe even pick up a few things. But what about robots that can truly understand how to interact with objects and their environment just by watching humans? That’s where HDMI or humanoid imitation for interaction comes in. And it’s a groundbreaking framework that teaches robots whole body humanoid object interaction skills directly from regular human videos. Think about that for a second. No complex programming for every little task. You just show the robot what to do and it does it. And it’s a big deal because up until now getting robots to interact with objects using their whole body has been super tough.

We just don’t have enough specialized motion data to do it. And the real world is full of these unpredictable contact rich situations. So how does HDMI pull off this magic trick? Well, it works in a three stage pipeline that’s as clever as it is effective. First, it’s video to data. It takes the human video and extracts both the human and object movements. And then it retargets the information basically translating human movements to robot and it builds structured data sets. Second, the reinforcement learning policy training takes effect where it uses this RL to train the robot.

This is where the robot learns by trial and error. And with a clever boost, it uses a unified object representation to understand different objects. It’s a residual action space for fine tuning movements and a general interaction award that helps it learn how to succeed in various tasks. And third, it uses zero shot deployment, which means that it can just be deployed directly to the real world from simulation without any extra tweaking. Beyond the latest in AI control. Let’s take a look at the latest in robot hardware. This is the Kangaroo robot from power robotics, and it stands 160 centimeters tall or about five feet three inches and weighs 40 kilograms or about 88 pounds.

And this robot has 28 degrees of freedom throughout its body for more human like articulation and runs for about six hours on a single battery charge. But here’s potentially the most important part. It runs on Linux Ubuntu and ROS. That’s robot operating system, which means that it’s specifically designed for robot researchers. Now let’s take a look at the hardware. First, there’s leg length actuators, which is the motor that adjusts the physical length of the robot’s leg. And Powell says they decoupled kangaroos leg length from its hip swing motions. And this is a big deal for the robots battery efficiency because it means no wasted energy on unintentionally affecting other movements and thus needing to expend energy on correcting the motions.

But even more importantly, it means that it’s simpler and more precise for various AI control policies. So the robot doesn’t have to constantly calculate for complex interconnected movements, which makes it more straightforward for controlling dynamic actions like jumping and running. On top of that, pal has developed in house electronics for motor control and sensor acquisition. This means that the robot features closed loop force torque controllers at the joint level, meaning that each joint can precisely measure and control the forces and torques that it applies. And this provides a huge boost in algorithm flexibility because direct control and feedback are crucial for testing complex algorithms that need fine grained control over interaction with the environment like balancing impact absorption and object manipulation.

Then there’s the safety aspect and stability because the direct control makes it easier to develop algorithms for more aggressive or nuanced control strategies without potentially damaging the robot. This provides an operating cushion of forgiveness, so to say. And next, there’s a new game creation AI agent. It’s called Gambo.ai. And they say that they take months of coding and reduce it down to just one prompt to build an entire game with all assets, including the characters, maps, music, sound effects, and animation being AI generated and auto organized. On top of this, they say that they’re building in ads and a map editor, and it looks like they’re creating some kind of a marketplace on their website where people may be able to make money off of the games that they create.

And now for the latest in text to video and image to video generation, starting with the release of WAN 2.2. This is an upgraded foundational video generative model with the following four key aspects. Firstly, WAN 2.2 integrates a mixture of experts architecture with video diffusion models. These are specialized to handle denoising across time steps, which increases the model’s capacity without increasing its computational costs. Second are its aesthetics. WAN 2.2 is trained on curated data with labels for lighting, composition, contrast, and color tones, which results in more precise control when customizing video styles.

Third, it features enhanced motion generation, where it uses 65.6% more images and 83.2% more videos than WAN 2.1. And this improves the generalization in motions, semantics, and aesthetics, achieving top performance among open and closed source models. And fourth is its high definition hybrid text and image to video. The open source 5 billion parameter model generates 720p videos at 24 frames per second from both text and images, and the whole thing can run on an RTX 4090 GPU. And to streamline 3D creation, there’s the release of SpatialGen, a multi-view multimodal diffusion model that automates high fidelity 3D indoor scene generation, and it leverages a new synthetic data set of over 12,000 annotated scenes, over 57,000 rooms, and 4.7 million photorealistic 2D renderings to address bottlenecks in visual quality, diversity, semantic consistency, and user control.

And the features include layout guided synthesis, where it takes the 3D layout and text derived reference images input. It can then output appearance, geometry, and semantics from any viewpoint for increased spatial consistency across modalities. Next, it also incorporates multimodal diffusion, which is an architecture that balances realism and control to outperform prior methods. Then its data set is curated for indoor environments, which enables training for design, VR, and robotics applications. And finally, it’s an open source release with its data sets and models being made available to the community. And you can use it for free on hugging face right now with its early demo.

And last but not least, in the latest AI audio generation tools, there’s the release of Xiaomi’s MIMO Audio. It’s a seven billion parameter generative AI language model pre-trained on 100 million hours of data. And it unifies text audio processing for speech continuation, in context learning, understanding, and dialogue across domains. But check it out for yourself. Here’s some continuation sound bites. The world has opened its heart of light in the morning. Come out my heart with I love to meet it. My thoughts shimmer with these shadows that woo me in the evening. Today night in its dusk has brought with it a tale full of remembrances.

He returned to Salt Lake City, careless what became of his own life, as long as he obtained what he knew to be justice. There he found evil tidings awaiting him. There had been a schism among the chosen people. And even that start which he’s in health had helped create was breaking apart upon its foundations. Anyways, tell me which models you’d like to see me cover next. Like and subscribe for more of the latest in AI news. And thanks for watching. [tr:trw].