Gemini 4 ARGON AI Agents Supercharge Humanoids, Blender, 3D Printing, 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: Gemini 4 ARGON AI Agents Supercharge Humanoids, Blender, 3D Printing, Animation · Channel: AI News · 07:55 · Watch on YouTube

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

Summary

The video presents demos of Gemini 4 Argon, focusing on how agentic loops can support robotics, 3D design, prototyping, and animation. In one example, an Approniks Apollo 2 humanoid learns to play cornhole by analyzing video, simulating actions, and correcting its own mistakes. Other demos use multi-agent workflows to create Blender scenes, design 3D-printable furniture parts, and build a racing game from photos of a real desk. The video also shows Argon generating customizable animations, including music, entirely through code.

Key points

The Apollo 2 humanoid learned cornhole from a single video by watching, simulating, failing, and adjusting its actions.
Argon analyzed successful throws, tuned release speed, widened the robot’s stance, and adjusted its shoulder to avoid collisions.
Google demonstrated building physics simulations from robot recordings, human videos, and sketches.
A separate robot demo used a closed visual loop to arrange Scrabble tiles into the word “Argon.”
Argon generated a Japanese temple complex in Blender using a prompt, architectural references, and a multi-agent critique loop.
A furniture-design demo turned a daybed photo and handwritten measurements into design options and a 3D-printable model.
Argon created a playable racing game modeled on objects in photos of a real desk.
The animation demo generated visuals and music from code, with a graphical interface for real-time customization and sub-agents working in parallel.

Chapters

Start	Chapter	Summary
00:00	A humanoid learns cornhole	The Apollo 2 robot learns to throw bean bags by analyzing a video and refining its actions through simulation and physical tests.
02:09	Building simulations and controlling robots	The video outlines ways to build simulations from robot recordings, human video, or sketches, then shows a visual-loop tile-sorting demo.
03:45	Generating a Blender temple scene	Argon uses a prompt, reference images, and a multi-agent workflow to create and animate a detailed temple scene in Blender.
04:47	3D printing and a photo-based racing game	Argon designs printable furniture legs from a photo and measurements, then turns desk photos into a playable racing game.
06:20	Creating customizable animations with code	Argon generates animations and music without imported media, while a custom interface and parallel sub-agents support iteration.

Quotes

It finally happened. — 00:00

with about 4 1/2 m being the sweet spot. — 01:05

it didn’t use one static prompt to do it. — 03:13

The result is a playable miniature racing game — 05:49

without a single video, image, or sound file. — 06:20

Concepts and entities

Gemini 4 Argon, Apollo 2, Apptronik, Google DeepMind, Blender, Nano Banana, Antigravity, Gemini Omni Flash, Three.js

Full transcript (original language)

00:00 It finally happened. Google just released its newest Gemini 4 Argon with over a dozen demos showcasing the next frontier of AI for humanoids, 3D design, 3D printing, and more. So, today on AI News, we’re looking at the very best ones to find out what the next generation of agentic loops just unlocked. Starting with Approniks Apollo 2 humanoid robot, which just taught itself cornhole from a single video using a loop of watching, simulating, failing, and fixing. And the speed at which it learned was shocking because

00:34 its first throw missed by just 86 cm. But by the 18th throw, it was reliably sending bean bags one after the other into a 15 cm wide hole. And Google says it’s because Gemini 4 Argon was diagnosing each failure itself by doing something incredible, which you’re about to see in just a moment here. So, first, Argon noticed that the human’s winning throws didn’t actually aim at the hole. Instead, winning throws usually landed on the tilted wooden deck above and then

01:05 let friction slide the bag in. Next, Argon solved for release speed by finding out for itself that just under 4 m/s fell short, but just over 5 m overshot, with about 4 1/2 m being the sweet spot. Then, it fixed each of the physical problems one at a time. A fast swing could tip over the robot, so it widened its stance. Meanwhile, a straight swing could clip its thigh armor, so it ran its own 3D collision check and shifted its shoulder outwards

01:35 for more clearance. And on the first real throw, the bag slipped out early and Argon traced the cause through its wrist camera to a halfopen grip that couldn’t hold 9.4 Newtons of pull at the bottom of the swing. And Google says the robot then sank seven consecutive shots. And in its own narration, Argon counts six straight shots with zero thigh contact. Furthermore, Argon was able to make this possible by using real to sim to rebuild the world inside a physics simulator, practice its moves there first, then return with what it learned

02:09 to be executed on the physical robot with Google showing three ways in particular. First from a robot’s own recording. In this case, Argon watches a real teleoperation episode and builds a digital twin. It finds or creates simulation assets that behave like real objects when the hands touch them, then keeps tuning until the simulation matches the real footage, including the contact events. Second was from the human video, which is how the cornhole example worked. Argon picked out the objects that mattered, swap crude boxes

02:40 for detailed shapes, so containers kept their openings, and mapped the human’s hand motion onto a humanoid with 76° of freedom throughout and 22 in each hand. And the third way was from a sketch where in Google’s example, a developer snapped a photo of a rough pen and paper drawing of the humanoid throwing a ball and then Argon built the simulation, worked out a balanced full body throw, and then verified in the physics that all the boxes would fall. But Argon doesn’t always need a physics simulator because in another demo, it steered a

03:13 humanoid with Gemini’s omni world model to rearrange five scattered scrabble tiles to spell Argon. The twist is that it didn’t use one static prompt to do it. Instead, Argon worked in a closed visual loop where it wrote timestamped action prompts, inspected key frames from the models rollouts, and then refined its prompts. But controlling robots is just the beginning of what Argon showed off. Because in a similar loop of looking, critiquing, and fixing, it also demonstrated breakthrough results in 3D modeling from a single

03:45 prompt. To start, Google’s DeepMind team pointed Gemini 4 Argon at a Japanese temple complex in Blender, and the whole thing came from just one prompt. And 3D builds are a tough test because it’s a complex coding task where mistakes are instantly visible. So, the team has been using pagotas as a test subject since last year to track their progress. And this time, they wanted photorealism built through Blender’s headless command line. But the prompt was far more than just build a pigota. And from there, the prompt points to a skill, which is

04:17 documentation the model reads for building large coherent 3D scenes. Then it has Gemini call a tool that creates architectural blueprints and has Nano Banana turn those into four reference images of how the finished temple should look. Next comes the main event where a multi-agent adversarial loop splits up the work between different agents with an orchestrator overseeing everything and an adversarial critic hunting for flaws. While Blender’s renders keep getting compared against those nano banana images, then the team came back

04:47 the next morning to see the finished renders, which the model had animated and generated bell sounds for, as well as ambient effects. But what happens when you point Argon at something that you can actually hold in the physical world? Enter the world of 3D printing and prototyping, where in this demo, Google gave Argon a photo of a daybed, a sticky note with handdrawn measurements, and a simple request to make a 3D design app that explores 10 different options for funky new legs. That was the start of it, just with a photo and a note. So

05:18 working through anti-gravity in the Gemini Mac OS app, Argon used Nano Banana to generate design iterations, then produced a 3D printable file for one of them, building the model to the correct size to fit the actual sofa. But Argon can also mix the physical world with digital generation because in another demo, a creator took a few photos of their messy desk, handed them to Gemini for Argon, and simply asked it to build a racing game based on them. So to start, Google says Argon ran a multi-agent loop in anti-gravity,

05:49 similar to the temple build, spinning up sub aents that used Blender to model the desk objects, the room, and the racetrack. Those models were then collision checked so the obstacles sat exactly where the real objects were on the desk. And from there, the scene was textured with Nano Banana, then wired up with physics, custom shaders, and car engine audio from Gemini Omni Flash, all running in 3JS in one build. And the result is a playable miniature racing game where you drift around the actual coffee mug and launch off a ramp over

06:20 your real keyboard. Both modeled from the real ones in the photos. But even that started with photos. So what happens when Argon has nothing to start from? It turns out it can still create because Google also showed it making entire animations without a single video, image, or sound file. So here, Google’s creative team used Gemini 4 Argon inside of anti-gravity to build animations using only code. So everything, including the music, was generated live with nothing being imported. And Argon didn’t just produce

06:50 beautiful results. It actually understood their intent well enough to generate different variations whenever they asked. So at first, they took an exploratory approach, testing art styles they already knew, and Argon generated a version of each. But instead of prompting every single change themselves, they asked Argon to build a graphical interface for customizing the animation and the music in real time, which let them tweak each element by hand until it looked exactly how they wanted. And on top of that, Antiggravity’s sub agents can tackle multiple tasks in parallel. So in the

07:23 time it takes to make one version, they say you could end up with 10, 20, or as many as you need. So that’s Gemini 4 Argon. Make sure to comment down below which demo was the most impressive. Whether that was the robot cornhole, 3D architecture, 3D printing, or something else, and which of these would you probably use first, or would you do something else? Anyways, like and subscribe for more of the latest in AI robotics news, and

07:55 check out this video here. And thanks for watching.