New NVIDIA Powered Gen 4 AI Robot, New IRON AI Brain (CLAUDE DUET)

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: New NVIDIA Powered Gen 4 AI Robot, New IRON AI Brain (CLAUDE DUET) · Channel: AI News · 08:02 · Watch on YouTube

Topics: AI agents, Task and workflow automation, Multimodal models, Robotics and autonomous systems

Summary

The video surveys developments in humanoid robotics and AI agents, beginning with Neura’s Gen4, a lighter robot with responsive limbs and locally running Neuron AI that learns from real-world training data. It then covers Xpeng’s upgraded VLA 2.0, designed to improve robots’ real-time perception, memory, and prediction, and ETH Zurich’s contact-planning system for humanoid mobility. Nvidia’s Flora data engine is presented as a tool for labeling robot-training video at scale, with performance and cost claims noted as company projections. The video concludes with Anthropic’s Claude duet, a proactive agent that monitors connected apps and requests approval for consequential actions.

Key points

Neura says its Gen4 is 20% lighter than its predecessor and features advanced hands, reflexive smart limbs, and a soft textile exterior.
Gen4 runs Neuron AI locally, combining vision, touch, force, sound, and the robot’s state; skills can be shared with other connected robots through the neuroverse.
Neura’s Robo gym in Germany trains humanoids on real-world tasks, and Gen4 is being prepared for series production, with no release date or price announced.
Xpeng’s VLA 2.0 maps visual signals directly to actions, and its Infini-VLA upgrade adds up to 30 seconds of memory and a predictive model that simulates up to 6 seconds ahead.
Xpeng says its next-generation Iron robot has 82 degrees of freedom and is targeting large-scale production by the end of 2026.
ETH Zurich’s ACP system plans contacts for nine body parts, enabling a Unitree G1 to climb, vault, and crawl; the reported success rate was 92.5% on eight obstacle courses.
Nvidia’s Flora generates video captions, progress ratings, and success or failure judgments for robot training and monitoring.
Nvidia claims Flora labels 200,000 hours of video per day and costs about 72 cents per video hour, while noting that comparisons with Astra are not like-for-like.
Claude duet sends morning briefs and monitors connected apps, while asking for approval before actions such as sending messages, making payments, or deleting items.

Chapters

Start	Chapter	Summary
00:00	Neura unveils the Gen4 humanoid	The video introduces Gen4’s lighter design, responsive limbs, locally running Neuron AI, training gym, and production status.
02:10	Xpeng upgrades its humanoid AI	Xpeng’s VLA 2.0 and Infini-VLA are described, including their visual-to-action approach, memory, prediction, hardware, and plans for the Iron robot.
04:17	ETH Zurich’s contact-based mobility system	Researchers use attention-based contact planning to help a Unitree G1 traverse obstacles, reporting a 92.5% success rate across eight courses.
05:21	Nvidia Flora and robot-training data	Nvidia’s Flora labels robot video at scale, with the video summarizing the company’s cost and benchmark claims.
06:26	Anthropic Claude duet	Claude duet monitors connected apps and sends briefings or alerts, while requiring approval for consequential actions.

Quotes

What if the real race isn’t who builds the best humanoid, but who teaches it fastest? — 00:00

It all runs locally on the robot. — 01:03

the model now remember up to 30 seconds of what just happened — 03:11

the harder problem is labeling thousands of hours of video. — 05:21

it even shows you an action and waits for approval — 06:57

Concepts and entities

Neura, Gen4, Neuron AI, Robo gym, Xpeng, VLA 2.0, Infini-VLA, IRON, ETH Zurich, ACP, Unitree G1, Flora, NVIDIA, Claude Duet, Anthropic

Full transcript (original language)

00:00 What if the real race isn’t who builds the best humanoid, but who teaches it fastest? Today on AI news, we’re taking a first look at a brand new humanoid robot, a new brain with new skills, plus the next frontier of cloud AI agents. But first, Neura just unveiled the next generation of its humanoid, the Gen4, and it’s already working. So here’s what’s new, broken down into three parts, starting with what you can see, which is the physical design. So Gen4 is

00:30 slimmer and 20% lighter than the last generation with a soft black textile skin to make it look more approachable. Plus the lighter body enables Gen4 to use its second upgrade even more effectively, which are its hands. Specifically, Neura says Gen4 has advanced hands for more accurate grips, plus what it calls smart limbs with reflexes built into each limb, allowing it to react on its own and move more fluidly. But once limbs can react, something has to guide them, right? So that’s the third upgrade, which is

01:03 Gen4’s brain, and Neura calls it Neuron AI. It feeds vision, touch, force, sound, and the robot’s own state into one model, which all help it understand its environment, and it all runs locally on the robot. Then, the intelligence distributes itself through what Neura calls the neuroverse, so that a skill learned by one robot also becomes available to other connected robots. But where they learn these skills is another key, which is Neura’s recently opened Robo gym in Germany, where companies teach humanoid robots real-world tasks.

01:38 In fact, companies from more than 10 separate industries, including automotive, pharma, and logistics, have all been training there since September, and Neura says Gen4 is learning there, too, from real-world robot data, human demonstrations, and simulation. And as for release, Neura says for anyone, Gen4 is being prepared for series production, but there’s no specific date yet or price range. So, how much should this cost in comparison to other humanoid competitors, and where do you think it ranks on the totem pole? Because Xpeng

02:10 also just revealed its newest humanoid upgrade, but from an angle that’s all about the next frontier of real-time perception and reaction. And Xpeng says the brain behind it is an upgraded VLA 2.0 model that now runs on the robot itself. But to see what’s new, you have to start with what VLA 2.0 was before. So, first last November, Xpeng launched VLA 2.0 with a different path. As most vision language action models turned what they see into language before

02:41 acting, but Xpeng said its model used a vision-implicit token action path that cut the language step out, going straight from visual signals to actions. And Xpeng also called it the first mass-produced physical world model, and said it could drive cars, humanoids, and flying cars being trained on nearly 100 million video clips without manual labeling with a 72 billion parameter base model in the cloud. So, enter the upgrade. According to Xpeng, the new

03:11 architecture called Infini-VLA lets the model now remember up to 30 seconds of what just happened, and a predictive world model called X foresight simulates up to 6 seconds ahead, though the hardware could reach 21 seconds, and Xpeng caps it to save compute. In fact, they also say the on-device model has 3.5 times the parameters of the last one with response speed up about 300%. So, the car’s already got this at Xpeng, rolling out from late August on models

03:43 with three Turing chips, each rated at 750 tera operations per second with a light version for its single-chip cars following in September. And as for the iron robot, the upgrade gives it stronger temporal understanding and future behavior prediction. So, instead of recognizing a gesture once and finishing a routine, it keeps observing and adjusting in real time. With Xpeng saying its next-gen iron has 82° of freedom, including a 22° of freedom hand, and uses three Turing chips for 3,000 tera operations per second. With

04:17 Xpeng aiming for large-scale production by the end of 2026, starting in its own stores. But, a robot that can shake your hand still has to get to you, and that’s a different problem. Which is why researchers at ETH Zurich just revealed ACP, or humanoid mobility via attention-based contacts. And the idea is to plan where the body touches the world instead of copy whole motions. So, first, the robot is a Unitree G1 with 23 joints and a single policy handles

04:48 climbing ladders, vaulting walls, and crawling. And it works by having one network plan the next contact for nine body parts, including hands, elbows, knees, feet, and pelvis, up to 1.5 seconds ahead, using a height map from onboard lidar. And a second [music] controls the joints. And here’s what’s remarkable. The contacts were learned from just 13.2 minutes of human motion, with reinforcement learning filling in the rest. And the results were astounding because on eight composite obstacle courses, the authors reported a

05:21 92.5% success rate versus just 58.8% for earlier methods. But, while ACP learned to climb from just 13 minutes of human motion, at robot company scale, the harder problem is labeling thousands of hours of video. That’s the job of Flora, a new data engine from Nvidia’s GEAR Lab, used to train its Groot robot models. In fact, Nvidia says it produces timestamp captions to filter training data, step-by-step progress ratings for

05:51 reinforcement learning, and success or failure judgements to monitor robots in the real world. And it works at scale, labeling 200,000 hours of video a day across 512 GB300 GPUs, with Nvidia claiming that it’s roughly 125 times cheaper than GPT-6 Astra based on about 72 cents per hour of video against Astra’s $88. But those are projections, not like-for-like measurements. And on quality, Nvidia says Flora beats Astra on some benchmarks, including rewards on

06:26 Robo reward, though Astra still leads on others, like accuracy on WGO bench and Robo meter. So, Flora helps monitor what robots are doing, but what about keeping track of everything that you need to do? For that, Anthropic just launched Claude duet, which is its newest app to make Claude work as a proactive agent. In fact, Claude duet simply sends a short morning brief across your connected apps, watches them through the day, and messages you only when something needs attention. And it runs on its own

06:57 virtual machine with its own browser, so your phone or laptop doesn’t need to be on. Plus, Anthropic says it even shows you an action and waits for approval before sending messages, making payments, signing documents, or deleting anything. And when it comes to access, currently there’s a waitlist starting with US Pro and Max subscribers on iOS, Android, and claude.ai, with further availability to be determined. Anyways, like and subscribe, and check out this video here for more of the latest in AI and robotics news,

07:30 and thanks for watching.

08:02 >> Woo!