Tesla Killer IRON 82 Joint Humanoid Robot Hardware Revelation (AI NEWS)

And this gets us into the muscle design. Here you see that they’re using a hollow 3D lattice structure and this allows for them to recreate the same compression and elasticity of human muscles. But there was a challenge because they had to train custom reinforcement learning algorithms to be able to understand how these 3D lattice muscles would affect the robot’s dynamics, which leads us to the iron’s body proportions. So the team was aiming for the golden ratio of the human form factor but conventional actuators were too bulky and so instead they built these custom smaller lighter actuators that you can see here to match the actual human muscle structure.

And because of that they’re able to achieve this extremely human-like gait. But there’s more than just the way the robot walks because there’s also the skin and it’s not just a shell. Here you can see that they were actually using these different materials to be able to fit around the form of the robot and sync with its lattice muscles underneath. And these actually cover up an entire built-in touch sensor suite which allows for the robot to have an even greater level of understanding in terms of its hardware. But we’ll get into that in just a second.

And you can see the way that these 3D lattice structures flex. So beyond just the skin and the muscles there’s intelligence and power. We have a 3D curved face display but what’s most important here to control the robot is its three touring AI chips with 2,250 tera operations per second to allow it to achieve this gait. And it also features three onboard LLMs which include a new vision language task model that acts as the core brain of this robot. Plus, what could be the most impressive in terms of just raw tech is this solid state battery which is an industry first according to XPong.

Fits right in with a 30% lighter footprint but 30% more power and it can operate safely at up to 250 Celsius while withstanding 300 grams of force or even three millimeter deep punctures. So safe to say this robot should be able to work in the real world and take a significant amount of wear and tear. But just how much of the data produced by the robot is going back to XPong? What level of privacy is there? That’d be interesting to know. And in terms of XPong’s deployment, they’re going for an open SDK for global developers to be able to build on top of this robot and they’re aiming to compete with Tesla at a $20,000 to $30,000 price range.

But the real question here is what can this robot do in the real world? We see some clips of its dexterous five-fingered hands but comment down below what you would want to see. What would it take for this robot to be genuinely useful to warrant its $20,000 to $30,000 price tag? And speaking of robots learning to move like humans, NVIDIA researchers just dropped EgoScale, a framework that tackles this exact problem but from a completely different angle. Instead of programming robot movements by hand, they’re training robots by watching us. And the team has already built a vision language action model trained on over 20,000 hours of egocentric human video, which is just footage from a human’s perspective, making it now 20 times larger than any previous effort.

And they’ve discovered something fascinating, which is that there’s a predictable log linear scaling law between how much human data you feed the system and how well it performs, meaning that more data equals better robots on a more reliable curve. And the transfer recipe is surprisingly simple. Massive human video pre-training combined with a small amount of aligned human robot mid-training, and the result is a 54% improvement in success rate over baseline using just a 22 degree of freedom robotic hand, which is notably the same amount of degrees of freedom as iron’s hand.

And on top of this, the learn motor skills transfer across different robot hand designs, meaning that the human motion data acts as a reusable embodiment agnostic motor prior. So whether it’s a 22 degree of freedom dexterous hand, or just a simple gripper, the foundation holds. So this leads to a future where robots will be able to generate the physics needed to move and do different tasks. And speaking of generative AI, anthropic just released Claude co-work desktop, which now lets users schedule tasks to run automatically, like morning briefings, weekly spreadsheet updates, or presentations.

And new plugins add specialized knowledge as well. Notably, it’s available on Mac OS and Windows for paying subscribers and as with any AI agent. But there’s another breakthrough in generative AI as Google just dropped nano banana to combining pro level quality with high speed. And it features advanced world knowledge via web search precision text rendering as well as translation plus subject consistency across up to five characters and 14 objects with outputs ranging from 512 megapixels to 4k. Overall nano banana two is delivering sharper details, richer textures and better instruction following.

And it’s all available now alongside nano banana pro with synth ID built in. Anyways, like and subscribe check out more of the latest in AI news here and thanks for watching. Bye
[tr:trw].