tech

What We Know About OpenAI's GPT-6 Sol and the Microduck Robot

News

· tech, science

A small white robot on four legs standing in a living room facing a bengal cat
Photo by Sanket Mishra / Pexels · source

A WIRED reporter says OpenAI's GPT-6 Sol model let him train a virtual version of Microduck, a small robot from Pollen Robotics and Hugging Face, to walk around a room and identify a bengal cat in a few hours, work he says previously "might have taken me weeks," according to WIRED. The physical Microduck has not shipped yet, and the reporter is still waiting on a unit, per the same account.

What did the reporter actually build?

The project is a simulation, not a finished product. Using GPT-6 Sol, the reporter trained a virtual Microduck to walk without falling over and to spot a bengal cat named Leona in its field of view, WIRED reports. The eventual goal, described in the piece, is to have the robot bow its head to release a treat from a 3D-printed, head-mounted storage chamber once it locates the cat. None of that hardware exists yet; the behavior was developed entirely in simulation on the premise that it will "transfer over to the real thing" once a physical unit arrives, according to WIRED.

An earlier attempt to teach the virtual robot to moonwalk, using a GIF of Michael Jackson as a reference, is also described in the piece, though WIRED's account of that attempt is incomplete in the available excerpt.

How did GPT-6 Sol speed up the work?

The model did more than generate code, per WIRED. It used Nvidia's Isaac Sim software to send instructions to a model controlling the virtual robot, then watched video of the robot in action to judge how well a behavior was working and suggested changes to the training process. That loop — instruct, observe, adjust — is what the reporter credits for compressing a multi-week task into a few hours, according to the WIRED account.

What did UC Berkeley researchers find with a related OpenAI model?

Separately, WIRED reports that UC Berkeley professor Ken Goldberg and his students tested OpenAI's Astra model, released in September, after seeing its ability to render 3D images. A member of Goldberg's lab gave Astra access to Nvidia's Newton tool, showed it a video clip of a real robot, and asked it to build a matching simulation, per WIRED.

"What blew us all away was that within an hour it had produced a full 3D simulation with all the parameters set — the position of the camera, the lighting, the background, all the materials," Goldberg told WIRED.

Goldberg also said Astra can automate much of the repetitive work of training a robot in simulation by running experiments and reviewing results itself, according to WIRED — the same kind of automation the reporter describes experiencing with GPT-6 Sol on the Microduck project.

What hardware and software ran the simulation?

The reporter's setup combined GPT-6 Sol with an agentic harness called Hermes from Nous Research and an Nvidia DGX Spark, a personal computer built around a dedicated AI chip, WIRED reports. Isaac Sim, Nvidia's robotics-simulation platform, handled the physical modeling that let the virtual Microduck's movements and camera view be tested before any hardware existed. Goldberg's lab used a separate Nvidia tool, Newton, for its Astra-based simulation work, according to WIRED — underscoring how much of this round of robotics progress is routed through Nvidia's simulation stack rather than physical prototypes.

When does a physical Microduck ship, and what should readers watch for?

No ship date for the physical Microduck appears in WIRED's reporting; the piece notes only that the reporter is "still waiting patiently" for a unit from Pollen Robotics and Hugging Face. The timeline so far, as reported by WIRED:

  • September 2026: OpenAI releases Astra; Goldberg's UC Berkeley lab begins testing its simulation capabilities with Nvidia's Newton tool.
  • Within one hour of that test: Astra produces a full 3D simulation with camera position, lighting, background, and materials set, per Goldberg's account to WIRED.
  • Nov. 11, 2026: WIRED publishes the reporter's account of training a virtual Microduck with GPT-6 Sol, Nous Research's Hermes harness, and an Nvidia DGX Spark.

What to watch from here, based on what WIRED has reported:

  • Whether Pollen Robotics and Hugging Face set a shipping date for a physical Microduck unit.
  • Whether behaviors trained in simulation — walking, cat detection, treat dispensing — transfer to physical hardware without retraining, the open question the reporter flags in the piece.
  • Whether Goldberg's lab publishes results from further Astra-based simulation tests beyond the single demo described to WIRED.

For the full account, including video of the virtual Microduck in motion, see WIRED's original story.

Disclosure. This article may include affiliate links; we may earn a commission at no extra cost to you. Legal entity: Pinewood Creations LLC. Smorgi Apps appears only as an affiliate partner in house slots — not as publisher or owner. See our affiliate disclosure.

Questions

What is Microduck?

Microduck is a small robot developed by Pollen Robotics and Hugging Face; a WIRED reporter trained a virtual version of it before a physical unit had shipped, according to WIRED.

How did OpenAI's GPT-6 Sol help train the robot?

GPT-6 Sol used Nvidia's Isaac Sim to control the virtual robot, watch video of its behavior, and suggest training adjustments, cutting a task that could take weeks down to a few hours, per WIRED.

What did UC Berkeley find when testing OpenAI's Astra model?

Professor Ken Goldberg's lab said Astra produced a full 3D simulation with camera, lighting, and material settings within an hour, using Nvidia's Newton tool, according to WIRED.

Sources

More from HTT News

Briefing

Top stories from the HTT News network by email. Free. No noise.