Neuron AI, Boston Dynamics New ATLAS Factory, New DeepSeek AI, Argon 3D Animation

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: Neuron AI, Boston Dynamics New ATLAS Factory, New DeepSeek AI, Argon 3D Animation · Channel: AI News · 07:30 · Watch on YouTube

Topics: AI content generation, AI models, Robotics and autonomous systems

Summary

Neural Robotics presents Neuron AI as a physical-AI foundation model that combines vision with touch, force, audio, memory, robot data, and language, and says it runs locally on its robots. Boston Dynamics describes how its in-house machine shop and prototype operations team are speeding up Atlas development, with a single robot now taking about a day to build. The video also covers DeepSeek V4.1 Flash 1017’s reported coding benchmark results and low per-task cost, while noting that independent results are not yet available. Google showcases Gemini 4 Argon turning PDFs into interactive apps and generating a 3D room in Python, before the video closes with a robot that learned football through self-play in simulation.

Key points

Neuron AI is described as a perception language action model that supplements vision with touch, force, audio, memory, joint data, and language prompts.
Neuron says its model runs locally on robots, enabling real-time decisions without a cloud connection and keeping production data under the customer’s control.
Neuron says it trains using real-world robot data, simulated task variations, and human demonstrations recorded with sensor-equipped suits.
Boston Dynamics’ prototype operations team uses an in-house machine shop to make parts for robots that may not yet exist, including specialized gear cutting.
Boston Dynamics says building a robot has gone from a major event to about a day, with a goal of reducing that to around an hour next year and minutes in later years.
DeepSeek reports that V4.1 Flash 1017 scores 77% on the DeepSWE coding benchmark at just under 20 cents per task; independent results are not yet available.
Google says Gemini 4 Argon can turn dense PDFs into interactive animated web apps, with updates beginning to roll out on October 16 for AI Pro and Ultra subscribers.
Google says Argon generated a realistic 3D room from a short prompt using vanilla Python and no assets, writing 5,000 lines of math.
The closing segment describes a robot learning football through self-play in physics simulation for 140 years before being tested in a real-life match.

Chapters

Start	Chapter	Summary
00:00	Neuron AI and multimodal robot perception	Neuron AI combines vision with touch, force, audio, memory, and robot data, and Neuron outlines how it runs and is trained.
02:11	Boston Dynamics’ in-house Atlas production	Boston Dynamics explains how its machine shop and end-to-end collaboration are accelerating robot prototyping and assembly.
04:51	DeepSeek’s coding model and reported costs	The video presents DeepSeek V4.1 Flash 1017’s reported benchmark accuracy and low cost while noting the absence of independent results.
05:53	Gemini 4 Argon’s interactive apps and 3D room	Google showcases Argon transforming PDFs into interactive web apps and generating a detailed 3D room with Python.
06:56	A robot learns football through self-play	The final segment describes a robot trained against earlier versions of itself in simulation before playing a real-life match.

Quotes

vision alone isn’t enough. — 00:00

it runs locally on the robot — 01:07

now it takes about a day — 03:44

independent results aren’t in yet. — 05:22

Google says it wrote 5,000 lines of math. — 06:26

Concepts and entities

Neuron AI, Neural Robotics, Boston Dynamics, Atlas, DeepSeek V4.1 Flash 1017, DeepSWE, Hugging Face, ModelScope, Gemini 4 Argon, Google, AlphaGo

Full transcript (original language)

00:00 What if the next step for embodied AI isn’t better eyes, but everything else? Neural Robotics just published Neuron AI, which it calls the foundation model for physical AI, and its argument is that vision alone isn’t enough. Most models that control robots today are vision language action models, which Neuron says judge mostly by camera input and reason on the current frame. Neuron AI is what it calls a perception language action model, and the difference is everything it added on top of sight. But to see why that matters,

00:33 you have to look at what a camera can’t tell you. So first, it still sees, reading a scene as 3D space, including geometry, position, and how objects can be reached. Next, it feels, using touch and force data to sense contact, like whether a screw catches or strips, or whether a bag is full or empty. Then it hears, using audio to detect things out of sight, like a latch clicking into place. And finally, it remembers, tracking objects over time. So a part hidden behind another is still known to be there. Neuron’s own examples include

01:07 confirming a lock has opened by sound and touch alone, and judging weld quality from the sound of the process. All of that gets combined with the robot’s own joint data and language prompts into a single stream of actions. But a model doing all that raises a question, which is where it actually runs. And according to Neuron, it runs locally on the robot, so that decisions happen in real time with no round trip to a server, and it keeps working without a cloud connection, while production data stays under the

01:37 customer’s control, which the company presents as important for European manufacturers. Neuron also says the same model deploys across its robot family of four devices, including the 4 Any 1 humanoid. And to learn all of this, Neuron says it trains on three sources, with the first being real world data, where robots perform tasks in neuro gyms and real environments, recording vision, touch, force, audio, and movement together. Second is in simulation, where its data amplifier creates variations of tasks and environments, and third is

02:11 human demonstrations from people wearing sensor-equipped data suits. And while Nura is focusing on building a smarter robot brain, another company is taking a whole different approach. As inside Boston Dynamics, there’s a team called prototype operations, but their workspace looks much more like a full manufacturing line, because what the team just unlocked will bring Atlas to sales rooms much sooner than anticipated. And that is the ability to finally prototype at scale in-house, supporting research and development

02:42 builds from design all the way through maintenance. But the reason it all finally works has less to do with the robots themselves, and more to do with something they decided to build for themselves. And to understand it, you have to start with what happens every time they try to get parts made somewhere else. So first, they own their own machine shop because every time the team tried to get parts made outside, suppliers said they couldn’t do it, citing the part tolerances were too tight, which Boston Dynamics says is how they accelerate their pace of development. And it’s not normal

03:14 production either, because they’re machining parts for robots that don’t even exist yet, which means developing the machining methods, making the first pass at design for manufacturing, and choosing the right tools along the way. But there’s one process in that shop that stands out even more. That’s gear cutting, which the team says is very specific in the industry and rarely found elsewhere. So now, as soon as the team needs something, the team can find a way to make it, backed by new machines, state-of-the-art equipment,

03:44 and robot arms built into the assembly process. But what matters most is what all that equipment is doing to the pace. In fact, just 1 year ago, building a single robot was quite an event, but now it takes about a day, and the team expects that to drop down to around an hour next year, and down to just minutes in the years after. On top of this, the feedback loop is just as fast, with an idea on Monday that could be implemented by the end of the week across design, assembly, and even organizational processes. And once engineers see what

04:17 the shop can do, they start asking things that they otherwise wouldn’t, leading to the last piece here, which is the end-to-end ownership, where a machinist starts with the engineers to find the best approach, then develops the part from start to finish. And instead of making a part from an order with no idea where it goes, they can now see the part assembled the same day or the next, then walk upstairs to the humanoid department to watch it move. They also work directly with the engineers and can go back to them when a design doesn’t work. So, the question now is, if robot production has gone

04:51 from a day to an hour and heading to minutes, what does that mean for how quickly the next one will be designed? But speed isn’t only a hardware story, because DeepSeek just released a model that keeps getting cheaper per task. DeepSeek V4.1 Flash 1017 scores 77% on the DeepSWE coding benchmark at just under 20 cents a task, up from its predecessor’s 74.2% at around 50 cents. And on the same chart, GPT-3 6 Astra, Gemini 3.8 Flash,

05:22 and Claude Opus 5 run from $2 to over $10 per task, and none match its accuracy. The weights are open on Hugging Face and ModelScope, with DeepSeek saying it beats Claude Opus 5, GPT-3 5.6 Soul, Kimiko 3, and GLM 5.3 in several coding and agent categories. But these are DeepSeek’s own benchmarks, and independent results aren’t in yet. So, if agents keep getting this cheap, what do people build with them? Google has one answer, and it starts with a PDF.

05:53 Google shows canvas in Gemini using Gemini 4 Argon to turn dense PDFs into interactive animated web apps you can explore at your own pace. Instead of reading every page, you get something you can move through with the key moments brought to life, and Google says the updates are rolling out in the Gemini app starting October 16th for AI Pro and Ultra subscribers. So, if Argon can reshape a document that already exists, what can it make from nothing at all? Google gave it this test exactly. The question was whether Gemini 4 Argon

06:26 could render a realistic 3D room from scratch in vanilla Python, and from a short prompt with zero assets, Google says it wrote 5,000 lines of math. So, what does that look like? The result traces concrete walls, walnut shelves, soft shadows, and lake sunlight, and it even bends light through curved glass to focus through caustics on stone, the bright patterns that form when curved glass focuses light onto a surface. So, this is really a test of whether a model can work out the physics and the code together. And the twist is that,

06:56 according to Google, Argon made the video showing all of it. And finally, this robot used skilled AI’s intelligence to learn football by playing itself for 140 years. It was born in a physics simulation, and it played against increasingly difficult opponents, where each one was a previous version of itself. And once it was ready, it was challenged to a match in real life. And here were the results.

07:30 And AlphaGo taught them that self-play can lead to emergent superhuman behavior. So, comment down below whether or not you think you’d be able to beat this robot at soccer yourself. Anyways, like and subscribe and check out this video here for more of the latest in AI and robotics news and thanks for watching.