Persona AI Hyundai Welding Humanoid Robot with NASA Tech For 2027

Persona AI and HD Hyundai Robotics are joining forces to create humanoid robots for precision welding by 2027. These robots will be equipped with advanced AI and special tools to perform dangerous tasks in shipyards, reducing injuries and downtime. A new technology called Video Mimic will allow these robots to learn new skills by watching and replicating human actions from videos. This partnership is part of a larger trend towards using advanced AI and robotics to make workplaces safer and more efficient.

I’m AI News, and Persona AI just announced a partnership with HD Hyundai Robotics, which gives us a sneak peek into the cutting edge of real-world humanoid automation. So here’s the breakdown. Basically, Persona AI is teaming up with Vazil Company, an HD Korea shipbuilding and offshore engineering, and they’re planning to deploy humanoid robots for precision welding sometime in 2027. And this collaboration in particular was unveiled in Houston, Texas, United States, and it’s promising to transform shipyards into hubs of human-robot synergy. And here are the details of the partnership. Specifically, it’s a memorandum of understanding to create these welding humanoids, with Persona AI being responsible for developing the necessary embodied AI, where Vazil Company is building specialized welding tools and testing setups, and HD Korea shipbuilding and offshore engineering is going to handle the process of real-world deployment.

And then HD Hyundai Robotics will supply the AI training data for welding paths, with humanoid prototypes being scheduled for late 2026 and field tests and commercial launch being scheduled for 2027. Plus, Persona AI will endow these humanoids with extreme dexterity thanks to NASA-derived robotic hand, with the goal of reducing injuries and downtime in the shipyard. In fact, this partnership aims to build what they call smart shipyards, where humans and robots will work together, with humanoids performing the dirty and dangerous tasks using Persona AI’s modular platform to let them observe, reason, and act on the fly, and they’re looking beyond just shipyards as they hope to assist with labor shortages worldwide.

But wouldn’t it be cool if robots could simply watch videos of humans doing things, and then just replicate the actions in the real world to learn new skills? Well, now they can, as researchers from Berkeley just revealed a breakthrough they’re calling Video Mimic, and it’s a new pipeline to teach humanoids context-aware skills using casual video footage, and this lets them adapt to new situations and environments later on. But what stands out about Video Mimic’s reel-to-sim-to-reel pipeline is that it mines everyday monocular videos, and it turns them into training for humanoid robots.

And it accomplishes this by analyzing footage of humans moving around in real-world settings, and then it reconstructs both the human motion and the surrounding environment into four dimensions. Next, the data is retargeted to a human in a simulator where a single context-aware control policy is trained to mimic the actions. And the results are impressive as Video Mimic allows these robots to confidently ascend and descend various staircases with different step heights and angles, and in other examples the robots learn to sit and stand smoothly on chairs and benches, and they adjust to different object shapes and sizes, and the robots also quickly learn to navigate uneven ground, slopes, and even small obstacles with stable human-like locomotion.

And all these skills come from one policy, plus the system even shows potential for ego-view rendering, and this hints at the possibility for first-person navigation in the future. And this is important because it affects humanoid robot consumers by giving them a super cost-effective method to teach their robots diverse sets of skills for real-world environments, like inside their homes and offices, by simply giving them videos of them doing things. And this will translate to the acceleration of humanoids doing real-world assisted roles. And this is only the beginning because Meta is also laying the groundwork for advanced humanoid intelligence with three new AI models designed to make robots smarter, more collaborative, and more capable by interacting naturally with the real world, with the first of these being Meta’s Perception Language Model, or PLM for short.

And this is a vision language model trained on the largest human-labeled video dataset, plus it’s combined with synthetic data, and it’s available in 1, 3, and 8 billion parameter variants, performing best in fine-grained video understanding and spatiotemporal reasoning. And this just means that it’s able to understand and reason about the relationship between objects in space and time. And on top of this, Meta also introduced the PLM Video Bench, which is a new benchmark targeting complex visual tasks. And the second AI model is Meta’s Locate 3D, which lets you tell the robot something like grab the red cup on the table, and it instantly processes 3D point cloud data from its RGBD sensors.

Next, it uses its 3D encoder and decoder, and it pinpoints the object and then acts on it. And Meta is releasing its dataset of 130,000 language annotations across 1,346 scenes to double its existing data. And this is all part of Meta’s effort to pave the way for intuitive human-robot interactions. And this is where we get into its third AI model, which is Meta’s Collaborative Reasoner. And this is Meta’s framework for training AI agents to collaborate like humans by tackling tasks through multi-turn conversations. And basically, this evaluates social skills like communication and empathy by using goal-oriented tasks where agents debate and agree on solutions.

Then, a self-improvement method powered by a high-performance inference engine called Matrix proceeds to generate synthetic interaction data, and then it boosts the performance by up to 29.4% on math, science, and social reasoning tasks. And Meta is open-sourcing everything. And meanwhile, Google just unveiled the IO edition of its Gemini 2.5 Pro AI model, and this is a pre-release multi-modal model that’s built to handle text, images, video, and code with a particular strength for front-end development and complex programming. In fact, Gemini’s 2.5 Pro tops the WebDev Arena leaderboard, judged by humans for web app look and function.

And when it comes to performance, it achieves a score of 84.8% on the VideoMME benchmark showing off a demo of a video-to-learning app where it transforms a YouTube video into an interactive learning platform with a UI and structure. And it can also generate code visualizations from images, and Google says that developers can access Gemini 2.5 Pro via Google AI Studio or Vertex AI, with existing users having already been auto-upgraded. And with this, Google has now integrated Gemini 2.5 into Firebase, and there are some showcases of what it can do, starting off with a Melody Maestro.

This was built in under 10 minutes, and it’s a Music Tutor app that teaches piano and violin with lessons for all skill levels, and the AI just grabbed some sound libraries and then made some quick tweaks to the violin sound that were quite realistic. Next, there’s the Schedule Helper where in 30 minutes, this app syncs your Google Tasks and Calendar and then slots tasks around meetings, and it even guides developers with the OAuth app setup. And in another example, there was the Quantum Phenomena Explorer, where in three prompts, the app delivered an interactive quantum physics tutorial with the test, double-slit experiment, and more, and visuals to support it.

And there was Social Sync, which was made in 10 minutes, and it suggests prime posting times, organizes posts, and fixes clunky calendar layouts in seconds. Then there was Flowforge, where in 20 minutes, it let you build AI workflows using natural language, like using a weather API to plan outdoor adventures. Then the workout wizard was made in under 10 minutes, and it crafted custom workout plans based on someone’s fitness levels, muscles, and schedule. And last, there was Stonecaster, which was made in one prompt. This is a game that uses AI opponents in a dark mode UI, but Firebase is open to anyone to build custom apps just like these.

But like and subscribe if you want to know more about the latest in AI news, and check out this video here. [tr:trw].