Future AI Robot WALKER C Game Changer Humanoid Collab Tech ($41000 USD)

So this will probably either feature dexterous abilities right out of the box, or else there will probably be an upgraded tier with hands like Unitree’s G1 for about $70,000. But there’s a twist here, because just recently, Tiankong also released brand new footage of its newest dexterous hands completing tasks in the real world, which may very well be the same hands for this upcoming humanoid robot model that’ll be unveiled soon. And it’s also completing barcode verification using the robot’s computer vision system, as well as multi-scale invocation, which refers to a robot’s ability to quickly switch between and execute multiple complex tasks, either in sequence or simultaneously, and in this case, which is often in response to environmental cues or direct instructions from a human.

But in this case, the robot is scanning, boxing, sealing, and attaching labels for shipping. And while Ubtech is already achieving a degree of real world autonomy in terms of logistics, physical intelligence is pushing this envelope even further with several examples of next level autonomy using the AGI bot. And this is really important because unlike most AI specialists, which are just good at one thing, this VLA model is a generalist, meaning that it combines vision, language, and action into one single system, allowing it to see, understand, and do things in the real world across different robot embodiments.

So you can think of this as an AI having an entire cool box instead of just a single hammer. And then there’s the multi embodiment angle, because usually if you train an AI to run on a robot, it’ll be locked into that robot body forever, like a two finger gripper AI staying with a two finger gripper. But the AGI bots model can jump between robots with different setups, which is like going from a basic gripper to a multi fingered hand with no retraining needed. And that’s a big deal, because it means you don’t have to start from scratch every single time you swap hardware.

And the skills being showcased here are practical, and the robot is reacting in real time. And because it’s adjusting on the fly, it means that if a cup moves a bit, or if someone hands it something unexpected, it can adjust to this. And most robots would freeze up in real time, but this one has the human like knack for just rolling with a situation. So it means that this could be brought into real world homes, or businesses, and it could control complex multi fingered hands to do real world fine motor skills, according to whatever a human were to dictate to it, regardless of what it already knows.

And other models would need separate systems to join. And pies endgame year is to set up a unified model that can run on any robot with any task, which means that they’re going for the holy grail of AGI with this AGI bot, which is a single adaptable brain for robots. And it looks like they’re making strides towards pulling this off. But what if you could instead just step into a cockpit that lets you command a humanoid robot by just moving your own body. And this is exactly the vision behind HOMI, which combines a low cost exoskeleton with advanced reinforcement learning for a fresh approach to operating humanoid robots.

And HOMI stands for humanoid loco manipulation with isomorphic exoskeleton cockpit. And all that is to say that HOMI works by integrating an exoskeleton that features arm components, gloves and a pedal interface with the gloves allowing operators to manipulate its hands and the pedals controlling its lower body. And the result is a robot that’s capable of walking, squatting and maintaining stability even when handling significant external forces. And the HOMI system is trained in NVIDIA’s Isaac Jim environment, which allows it to develop policies for fast and reliable performance. But what’s even more impressive is that HOMI collects real world operational data and uses imitation learning to train autonomous behaviors, which can then be easily transferred back into simulation environments using Isaac simp.

And while HOMI is still in the research phase, its team is refining its capabilities as it’s eyeing real world applications in human to robot collaboration. But AI video generation is also getting crazy as runway just unveiled its Gen four video generation model. And Gen four just solved one of the biggest problems in AI video generation, which is keeping characters, objects and styles consistent across scenes. So runway says that Gen four now delivers a new standard in video generation with its secret apparently lying in its ability to anchor consistency in just one reference image paired with text prompts.

So whether it’s shifting lighting or its objects placed in wildly different settings, Gen four says that it can now keep everything recognizable and fluid. And beyond just aesthetics, Gen four is also using advanced physics simulation, which lets users tweak real world dynamics like a bouncing ball or water splashing. And the models rollout is targeting paying subscribers and enterprise clients first with plans to add reference enhancing updates soon after and the models rollout is going to first target paying subscribers and enterprise clients with plans to add reference enhancements soon. But Amazon also just up the ante with its newest Nova real 1.1 as it can now generate long format multi shot videos with improved quality and speed over its predecessor.

And in fact, with version 1.1, this model still excels at crafting six seconds single shot clips, but it adds its newest ability to stitch multiple shots into cohesive two minute videos, all while keeping a consistent style across the board. And Nova real 1.1 actually features a dual mode approach. In multi shot automated mode, you toss in a single prompt of up to 4,000 characters and let the AI weave a multi shot narrative for you. But if you prefer more control, you can use the multi shot manual mode or storyboard mode, which lets you craft each six second shot with its own prompt and even an optional 1280 by 720 reference image pulled from Amazon s3 or encoded from its base 64.

And it’s currently accessible via Amazon bedrock right now. But that’s not the only leap as Nvidia, Stanford, UCSD, UC Berkeley and UT Austin also just released a brand new way to transform how we create videos with test time training. And here’s how it works. The TTT layers are a game changer because unlike standard approaches, TTT hidden states are neural networks themselves packing a punch of expressiveness. And the researchers plug these layers into a pre trained transformer and unlock the ability to generate one minute videos from simple tech storyboards. And the results were stunning, as in human evaluations of 100 videos per method TTT outclass rivals like mamba to gated delta net and sliding window attention layers scoring 34 elo points higher for coherence and storytelling.

And it’s still not flawless, as there are still some glitches and artifacts likely because the base five billion parameter model isn’t robust enough. And efficiency also lags sometimes with resource limits having been capped at one minute. But the team is optimistic and they say that it’s just the start, as they believe that TTT is going to dramatically improve AI storytelling. [tr:trw].