Galaxy General Robotics AI Dream Control GO BIG Phantom RL Adam Grok 4 Fast FREE

➡ The article discusses three major advancements in AI and robotics: Any Two-Track, Dream Control, and Figures Go Big. Any Two-Track is a two-stage reinforcement learning setup that allows robots to follow any motion accurately, even in chaotic environments. Dream Control is an AI upgrade that enables humanoid robots to perform tasks more human-like without constant supervision. Figures Go Big is a large-scale data and learning model that aims to make home robots as smart as humans. The article also mentions other AI tools and robots, including PND botics’ Adam humanoid and Phantom’s reinforcement learning for locomotion.
➡ The app is fast, flexible, and allows users to modify it while using. It’s fun and interactive. Don’t forget to test it, share your thoughts, and stay updated with the latest AI news by liking, subscribing, and watching more videos.

Today on AI News, how far are we from being able to just talk to robots like we do to humans and have them execute tasks with an intuitive understanding of what to do and how to do it? Well, we may be closer than ever before with the latest three robot advances that I’m going to talk about today. Plus, keep watching until the end to see the latest AI tools in action. But first, we’re going to talk about any two-track. This is a general robotics fix for making humanoids track motions more reliably, no matter how wild or unpredictable things may get around them.

So think of it like giving your robot a super adaptable sixth sense for movement so that it doesn’t just fall over when it gets a curveball. See, the big headache with robots right now is that they need to nail a huge variety of tasks, but it’s extra hard for them when their environment is a total mess or when they’re randomly being shoved, for instance. So here’s where any two-track comes in. It’s a smart two-stage reinforcement learning setup that lets robots follow any motion faithfully, even when there’s total mayhem around. And the cool part is that it bakes adaptability right into the core of the executing actions instead of treating them like some bolted-on afterthought like normal.

There’s the base layer, which is a single all-in-one policy that’s cleverly designed to juggle all kinds of different movements without needing a bunch of custom tweaks for each one. And then there’s any adapter. This is the real-time hero where it peaks at the robot’s recent track record and tweaks things on the fly to close that pesky cinder-real gap, and it shrugs off things like bubbly grounds, and it avoids distractions like uneven terrains or surprise disturbances. And for the proof, they slapped it onto the Unitree’s G1 humanoid, and this is using zero-shot transfer, no extra fiddling needed, and it crushes motion tracking with all sorts of scenarios, even when it’s hit with multiple disruptions at once.

And it feels like a pretty solid leap, and all this just makes you wonder what kind of real-world robot task do you think that this could unlock next. And next in AI Robot Control is Dream Control. This is another AI upgrade to make humanoid robots way more human-like and useful. Just imagine your robot buddy could simply understand how to do what you wanted it to do without you having to babysit it. That’s the vibe. And after a full year of tinkering, they’ve cooked up this scalable setup that blends diffusion models. Those are the AI tricks that dream up realistic images or videos with reinforcement learning, and that’s the robot’s way of learning by messing up a ton and then getting better over time.

And the result of this is that the robot can pull off natural full-body moves in the real world, not stiff strutting across a lab floor. But what sets it all apart is the diffusion prior that’s pulled from human motion data. This guides the RL training so that the robot’s actions feel super intuitive and lifelike without needing a mountain of video from people remotely controlling robots. And it’s all trained in simulation. It deploys straight to hardware like the Unitree’s G1 running reel-to-sim with barely any sim-to-reel headaches, meaning that it doesn’t freak out when it hits the physical world.

And some of its skills include picking up and lifting stuff in the kitchen, opening drawers or doors without the awkward robot jam, nailing precise punches, kicks, or jumps, makes it more agile, almost like a Parker Pro, and handling two-handed tasks like sorting tools or mixing ingredients. And they even split up the workload in two parts. The RL decision happens right on the robot’s edge hardware for speed while the heavy generative AI part runs in the cloud. And the bottom line here is that this helps general purpose humanoids to adapt and interact like humans would without the hand holding that it would normally take.

And next, when it comes to acting just like a human, we have figures go big. This is figures Helix model. That’s their vision language action AI. And it just got two massive upgrades aimed at making home robots as clever as humans. And this is all through cranking up the data and learning scale. It’s almost like giving these bots a PhD in chaos. And here’s why it’s called go big. They’re building the biggest, most mixed up data set ever for humanoid training. And it’s hooked up with Brookfield who owns a ton of real estate.

It’s over a hundred thousand apartments, 500 million square feet in offices, 160 million square feet in warehouses. And they’re already filming in these spots with real people doing real things with an egocentric view for Helix to ingest. But here’s the magic. No special bot training is needed. Just these human clips can be fed in. And Helix here is go to the fridge, for instance, and then it just goes there. And what figures doing here is building on arms only gigs like doing laundry, dishes, or sorting packages to achieve full body action like walking, grabbing the works.

All this will enable speech to navigation and complex cluttered environments. A unified model handling both upper and lower body control for dexterous tasks and locomotion. And finally, the first ever end-to-end learning pipeline from human videos fed straight into robots, allowing anyone to just talk to the robot and give it commands. And then it gets interpreted directly to do it. Next, let’s take a look at the latest from PND botics as they showcase the latest walk from their Adam humanoid. Some of the stats from this robot include the fact that there are three different versions.

There’s the Adam U, the Adam SP, and the Adam light. And they share most of the same stats, except that, for instance, the Adam U weighs one kilogram more than the Adam light. And all of them have 340 newton meters of force, but the degrees of freedom differ with the Adam U only having 31 degrees of freedom, whereas the Adam SP has between 41 to 43 degrees of freedom. But here we can see already it’s walking with a quite natural human-like gait compared to other robots. And here we see the upper half of the robot body dexterous five fingered hands.

We don’t yet have an outer shell on the robot, which means that there are still various pinch points. For instance, you could catch your fingers between some of those joints or different environmental factors like a crud or dust could get caught in there, making maintenance a little bit more difficult. But once they slap an exterior body shell on top of this, it could be somewhere near ready to go for a commercial release. But tell us in the comments how much you think this thing will cost and when will it probably go on sale to the water market? Next, let’s take a look at the latest from Phantom.

The company says that they’re moving from model controller to reinforcement learning for locomotion. That’s how the robot walks around, said the policy is looking good and that they say it’s better than wheels. But tell us what you think. The robot here is taking some kicks from the founder from both the front and from the back. Although I do think that if you were to kick it maybe 50% harder, it could take a fall. Looking good so far, though. But XAI just released their latest model, that’s GROC4 Fast, and you can currently use it for free on open router.

This model is faster and cheaper than GROC4, but it comes with a few key features. Firstly, the fact that it uses 40% fewer thinking tokens than GROC4, which reduces its compute demands. Next, it leverages a unified model that handles both complex reasoning and quick non-reasoning tasks via its prompt steering to lower latency. And it performed very well on different benchmarks here, as you can see. And in terms of tool use, it supports web search, code execution, multi-hop browsing, and it’s trained with reinforcement learning. And you can try it for free here on open router, as well as on GROC.com, and its pricing is 20 cents per million input tokens.

Next, in the latest and coolest AI tools, this is CopilotKit using AGUI. So CopilotKit is an open source framework that lets you add AI copilots to your app. And it syncs the app’s UI, which is the front end, that’s the user interface, and the AI backend in real time, that’s the server. So users and AI can work together seamlessly. And it’s as easy to set up as copy-pasting this one single command. And it works on any AI model, including open AI, land graph, or custom AI. And it offers React components or APIs for custom user interfaces.

So AGUI is the protocol that’s powering CopilotKit. And it’s a simple open standard that streams these JSON events, which are messages or tool calls or state updates over HTTP to keep the front end and the backend in sync. And it ensures real time updates, efficient state changes, and works with any AI backend, solves issues like streaming, tool tracking, and concurrency. So together, it makes it dead simple to build apps where the AI and users can collaborate instantly with no lag or lock-in, and they can change the state of the app as they go.

It’s just more interactive and cooler. Anyways, make sure to try it for yourself and leave a comment down below as to what you think. Like and subscribe, and check out this video here for more in the latest AI news. [tr:trw].