Google one-shots humanoid training, Tesla gives robots skin, Faraday launches $2,490 humanoids, and more!

Google's Gemini 4 Argon just one-shotted robot training

One video, no teleoperation, no task-specific dataset, and no engineers hand-tuning reward functions. With just a short video clip of a human playing cornhole, Google's Gemini 4 Argon figured out the rest.

As in: trained a robot almost instantly. 1-shot to success.

That's the demo Google released alongside the Gemini 4 Argon launch this week. Argon watched the human throw, mapped the motion onto Apptronik's Apollo humanoid, built its own physics simulation in MuJoCo, practiced there in simulation, then moved to real hardware. After a few real attempts (failures), Apollo sank 7 consecutive bags straight into the hole.

Kinda mind-blowing.

"I'm still half in disbelief," wrote Norman Di Palo, who works on robotics at Google DeepMind.

Now, some context. 

This isn't a humanoid learning to clean your kitchen from a single YouTube video. Cornhole is a repeatable motion in a controlled environment, and Apollo still needed real-world practice. But the workflow is revolutionary: a frontier AI model effectively doing the work of an entire robotics engineering team.

And it wasn't just cornhole. 

Argon also taught a Boston Dynamics Spot to operate a foot-pedal trash bin. In another demonstration, it built a simulation from a single tabletop photo, generated more than 4,500 training episodes, and improved real-world task success from 56% to 75%.

The massive shift here is that the AI isn't just controlling a robot. It's designing the training process that makes the robot useful. We're seeing frontier models one-shot training for robots, which should give us smarter robots, and do it faster and cheaper.

And allow anyone who APIs into this tech to do it themselves.

Huge!

Standard Bots wants factory workers vibe-coding robots

Here's the other side of that robot-training revolution.

Standard Bots is pushing robot training out of the engineering department and onto the factory floor. Its new AI platform aims to let ordinary workers teach industrial robots using an iPad, without needing to understand robotics programming.

Think about what that means. Traditionally, introducing a new task to an industrial robot required a specialist, programming expertise, and often significant time and expense. Standard Bots wants to make that process accessible to the person who actually understands the work: the factory employee.

This is post-training democratization. A robot arrives with general capabilities, but the person who uses it teaches it exactly what to do in their particular environment.

Google and Standard Bots are approaching the same problem from different directions. Google is automating the robotics engineering process with frontier AI. Standard Bots is trying to put the final customization -- post training -- in the hands of ordinary workers.

Together, they point toward something pretty significant: a future where the people who need robots don't necessarily need to know how to program them.

Galbot's humanoid now runs 200 stores, and it's expanding to Hong Kong

Beijing-based Galbot's Xiaogai humanoid stands 173 centimeters tall, speaks Cantonese, Mandarin, and English, stocks shelves, processes payments, and occasionally dances for customers.

It also operates in more than 200 convenience stores across nearly 50 Chinese cities. And this week, 3 more opened in Hong Kong.

Each location is a 9-square-meter capsule, open 24 hours, operated by a single robot. No overnight clerk, no human staff on site. The Beijing flagship has reportedly been serving around 1,000 customers daily since August.

What's particularly interesting here isn't the robot. It's the deployment scale.

At 200 stores, Galbot is collecting something enormously valuable: operational experience. Thousands of customer interactions, endless variations in products and transactions, and a growing understanding of what goes wrong when robots run real businesses.

That's a very different proposition from a lab demonstration or a corporate pilot. It's also an opportunity for continuous improvement across an entire fleet.

The question now is whether the economics work once the novelty wears off. Because the novelty is part of the retail success story. If it does NOT wear off, Galbot may have found a remarkably scalable business model for humanoid robotics.

Agility heads for Nasdaq as Faraday Future launches a $2,490 humanoid

Two very different approaches to the humanoid business are heading toward public markets.

Agility Robotics, the company behind Digit, is preparing to list on Nasdaq through a $2.5 billion SPAC merger with Churchill Capital XI. Shareholders vote November 2, and the company expects to begin trading under the ticker AGLT following the deal's completion.

Agility has more than $300 million in contracted robot orders and commercial deployments with major customers including Toyota, GXO, and Schaeffler. If the transaction closes, it will become a rare opportunity for public-market investors to own a pure-play humanoid company with real commercial deployments.

Faraday Future is taking a different route.

The electric vehicle company is transferring its robotics business into FFR, a publicly traded company in which it will retain approximately 66% ownership. The deal values the robotics business at about $200 million, essentially giving Faraday a separately traded robotics operation.

And now Faraday has launched two remarkably inexpensive humanoids.

The FF Nova costs just $2,490. It's a 50-centimeter humanoid with 20 degrees of freedom, aimed primarily at developers, universities, and schools. There's also a $299 Nova Mini for children, with basic programming capabilities.

The first 100 Novas are invitation-only, so this isn't mass-market availability yet. And these are small educational and development robots, not competitors to Digit or full-size industrial humanoids.

Still, that's an interesting combination: access to public capital markets, an expanding robot portfolio, and humanoid hardware starting at prices approaching consumer electronics.

The humanoid market is clearly becoming a lot more diverse.

Tesla wants to give Optimus something like human skin

Tesla has filed a patent application for touch-sensitive robotic skin.

Published October 8, the application describes flexible tactile sensors that can be manufactured as flat sheets and then molded around three-dimensional shapes, including robot fingers and palms.

The idea is straightforward but important. A robot hand needs to do more than move its fingers accurately. It needs to feel the objects it's touching, measure pressure, and adjust its grip before something slips, breaks, or gets crushed.

Tesla's proposed approach uses conductive layers separated by a compressible material to measure pressure. More importantly, the manufacturing process relies on established techniques like printing, cutting, and thermoforming.

That could make tactile sensing easier to manufacture at scale.

There's a big caveat here of course ... this is a patent application, not a finished product. Tesla hasn't confirmed that the technology will appear in Optimus, and the filing doesn't establish its sensitivity or durability.

But it addresses one of the biggest remaining problems in humanoid robotics. Hands aren't just about dexterity. They're about sensing, control, and feedback.

And as we've discussed before, Tesla's robot hands are already a major manufacturing challenge. Making sophisticated tactile sensors cheaper and easier to manufacture could be an important piece of that puzzle.

Meet Codey, the humanoid that wants to fight loneliness

Not every humanoid needs to work in a factory.

At Stanford this week, a company called Mind Children showed off Codey, a humanoid specifically designed for social interaction. Codey has a relatively small body, realistic facial features, and the ability to carry on conversations, interact with people, and play games like Pictionary and hide-and-seek.

The company's ambition is to use robots in hospitals, senior care facilities, schools, and eventually private homes, particularly for people who live alone.

Rather than replacing physical labor, Codey is designed to provide interaction, engagement, and potentially companionship. Of course, there's a substantial difference between a robot that can hold a conversation and one that can meaningfully reduce loneliness. 

Codey is still a prototype, and Mind Children is targeting healthcare applications in 2027. But the company is betting that some of the most valuable things humanoid robots will do might have nothing to do with lifting boxes.

Number of the day: 10,000

That's how many robotic hands Chinese startup Linkerbot says it shipped in 2025.

And these aren't simple two-finger grippers. Its advanced hands offer more than 20 degrees of freedom, with prices ranging from $980 to $14,700. Linkerbot claims more than 80% market share in China, with customers including Samsung, Siemens, Stanford, and Tsinghua University.

The company closed a funding round in April at a $3 billion valuation. Now it's reportedly seeking a valuation of $6 billion in another round.

Founder Alex Zhou started Linkerbot in 2023 with a relatively simple observation: the humanoid industry was going to need a lot of hands.

And here's why that's interesting alongside Tesla's tactile-skin patent. We're seeing enormous investment in two complementary challenges: making robot hands dexterous and sensitive enough to perform useful work, and manufacturing them cheaply enough to put on millions of robots.

AI might be learning how to teach robots faster than ever. But those robots still need physical hardware capable of doing the job.

And apparently, the market for that hardware is already worth billions.

Sources

• HumanoidMegaHub — Google's Gemini 4 Argon teaches Apollo cornhole
• HumanoidMegaHub — Standard Bots hands post-training to factory workers
• HumanoidMegaHub — Galbot's humanoid runs 200 stores
• Agility Robotics — SEC registration declared effective
• HumanoidMegaHub — Faraday Future's robotics transaction
• The Next Web — Faraday Future's $2,490 humanoid
• Robotics & Automation News — Tesla's touch-sensitive robot skin patent
• ABC7 — Codey, the humanoid designed to tackle loneliness
• HumanoidMegaHub — Linkerbot ships 10,000 robot hands