Vibe coding robots: Standard Bots hands post-training to ordinary factory workers with iPads

So many recent innovations in robot training

Robot training is getting cheap, fast. 

This week alone, Google's Gemini 4 Argon taught Apptronik's Apollo to play cornhole from one video of a human. And today Standard Bots launched Standard Bots AI, which lets a factory worker build and change robot programs by talking to the robot.

Those sound like very different things, but they're very related.

Both are attacks on the same bottleneck: the expensive, slow, expert-heavy work of getting a capable robot to do one specific job in one specific place.

“If you can … allow a normal worker to post-train an AI model for their task, you can now automate things you just couldn't automate before,” Standard Bots CEO Evan Beard told me in an interview.

Three skills, one goal

Standard Bots, which raised $200 million at a $1 billion valuation in June, makes AI-native robot arms and is working on industrial humanoids. The new platform ships three skills on all of its arms:

• Agent: a chat interface that builds robot routines from a prompt, which Beard compares to vibe coding with Claude Code or Codex
• Locate: tap an object on the robot's camera feed, or describe it, and the robot finds, tracks and moves around it, with no hand-set waypoints or fixtures
• Inspect: custom defect-detection vision models, built with Roboflow, trained on the customer's own parts

There's also a beta feature called Showpath that learns a task from a few demonstrations.

The company claims Locate cuts programming time by 90%. That's a company number, not an independent one. But the logic is sound: traditional industrial robots replay saved positions, so every change in a part, a box or a layout means reprogramming.

“The boxes come down, they're normally-- they're green 'cause it's some holiday, and Amazon's running its special,” Beard says. A classic robot fails there. The goal, he says, is a robot that “can almost think and adapt to how things are.”

The training stack is being automated from both ends

Zoom out and there's a clear pattern across the industry.

At the top of the stack, frontier models and agentic platforms are automating the work of robotics engineers. Argon built its own MuJoCo simulation, ran system identification and corrected its own aim. In another Google example, it generated more than 4,500 simulated episodes from a single photo and lifted real-world task success from 56% to 75%. 

At the same time, Nvidia-backed General Robotics says its GRID platform cuts robot onboarding by up to 99.7% using agents that handle calibration, simulation and skill transfer.

In the data layer, the fight is over what's cheapest and good enough. Figure is spending more than a billion dollars on real-world human video. Antioch raised $32 million arguing that simulated data can do most of the work. Nature Will just bet on 1 million hours of first-person video.

Standard Bots is working the other end: the last mile, where a general capability meets a specific factory with specific parts. Today that last mile mostly belongs to systems integrators. Beard says installation “often costs more than the robot itself,” and his head of applied AI, Leif Jentoft, puts typical integrator projects at a million dollars.

If the expert work at the top gets automated by AI, and the last mile gets handed to the people who actually run the line, the integrator-shaped middle shrinks a lot.

Why this matters for humanoids

Standard Bots sells arms, not humanoids. But the problem is identical.

Every humanoid robot company pitching general-purpose labor has to answer the same question: who teaches the robot this particular job, at this particular site, and what does it cost? Right now the honest answer is usually the robot maker's own engineers, often with teleoperation. That doesn't scale to thousands of customers.

Beard frames it in adoption terms. The physical world's ChatGPT moment, he says, isn't only about capability. It also needs “the steep part of the adoption curve,” and that requires robots “accessible to a normal person to be able to unbox it and train an AI model.”

His stated ambition is big: “We wanna build the Anthropic of the physical world.”

That's a stretch for a robot arm company today. But the direction is right. The companies that make post-training something a customer can do themselves, in minutes, will be the ones that scale.

Standard Bots AI is available now.

Sources

• Standard Bots press release, Oct. 15, 2026 (via email)
• Evan Beard interview, Humanoid Daily podcast
• One-shot robot training: Google's Gemini 4 Argon taught Apptronik's Apollo to play cornhole(HumanoidMegaHub)
• General foundation model for physical AI? Nature Will bets on 1 million hours of video (HumanoidMegaHub)
• Nvidia-backed robot platform says AI wrote 80% of its code, cuts robot training time 99% (Forbes)
• Here's a $32 million bet that robots don't need a billion dollars of real-world data (Forbes)
• Standard Bots raises $200 million (Forbes)