Will Knight wants a robot to feed his cat. The WIRED senior writer, who covers artificial intelligence, has a bengal called Leona and a plan: a small walking robot that finds her, bows its head and drops a treat from a compartment printed to sit on top of it. The treat is not the news. The news is that the robot already works in software, and that an AI model did most of the training a robotics engineer would normally do by hand.

The robot is a virtual Microduck, the duck-shaped biped from Pollen Robotics, the robotics arm of Hugging Face. Knight taught it to walk around a room without toppling over and to pick out any bengal cat nearby, and he says the job took a few hours where it once might have taken him weeks. His column calls that the future of robotics. It is half right. The training loop has collapsed. What sits on either side of the loop has not moved much at all.

What the model actually did

Knight paired OpenAI‘s GPT-6 Sol with Hermes, an agentic harness from Nous Research, running on an Nvidia DGX Spark, a desktop machine built around a powerful AI chip. The model did more than write code. Through Nvidia’s Isaac Sim it sent instructions to the model controlling the simulated duck, watched video of the duck in action, judged how well things were going and suggested changes to the training regime. Run, watch, tweak, repeat: that is the slow, unglamorous part of robot learning, and it is the part that was automated.

Professional labs report the same shift. At UC Berkeley, professor Ken Goldberg’s group gave OpenAI’s Astra model access to an Nvidia tool called Newton, showed it a clip of a real robot and asked it to rebuild the scene as a simulation. “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. Astra can also run the experiments and examine the results, he says, which takes over much of the drudgery of training a robot in simulation.

The method has a name now. ASPIRE, a framework whose authors include Goldberg and Nvidia’s Jim Fan, has a large language model write a robot’s control code, read the execution traces when it fails, repair the program and file the working fix in a skill library that later tasks can reuse. Fan’s lab at Nvidia put the v1.0 code on GitHub under an Apache 2.0 licence in mid-October, according to ToolNavs and RuntimeWire. One demo has robot arms learning to solve a Rubik’s Cube. “It picks it up, tries to twist it, but has the wrong angle,” Fan told WIRED. “Then it reasons for a very, very long time, but it eventually is able to solve it because it learns the skills over time.”

The skill that got cheap is rented

Here is the part the “vibe robotics” framing skips. The ASPIRE paper says plainly that its performance scales with the frontier model underneath. Running on Anthropic‘s Claude Opus 5.5, it reports 99% average success across all 80 LIBERO-Pro manipulation tasks, a self-reported figure that has not yet been reproduced independently. The project’s own list of limitations says the system depends on a frozen frontier model, and that it is unverified whether smaller models can sustain the same debugging loop.

Follow that through. A hobbyist’s duck and a Berkeley lab’s arm now learn about as well as the large model they can call. The engineering expertise that used to gate robotics has not disappeared. Much of it now sits inside whichever frontier model the builder can pay to use.

The duck is still in a queue

The other side of the loop is hardware, and Knight is still waiting for his. Pollen opened Microduck preorders at $399 in late August with a pre-Christmas delivery target. Business Insider reported that orders passed $2.6 million in the first 24 hours, citing Hugging Face co-founder Thomas Wolf, and the checkout page soon moved new orders to a four-to-six-month window. Gadget Flow reports the store now warns that current preorders won’t ship until after April 2027. Training in simulation is the intended path: Pollen’s own workflow has owners train behaviours on a virtual duck and export them to the physical robot. But a policy waiting for a body has not yet met a rug, a sofa leg or a cat that refuses to cooperate.

Knight is candid about the rest. The vision system on his duck, he writes, does not work well enough to be deployed in a real product. When he fed the model a GIF of Michael Jackson to teach it the moonwalk, hours of effort produced what he calls a very jaunty backward jog. ASPIRE’s authors list the same family of gaps for real deployment: reliable success detection, safe resets, safety monitoring and calibration. Each of those is about knowing what actually happened in a physical room, which is exactly what a simulator cannot tell you.

Where the money is pointed

None of this is slowing interest. Goldberg says a number of industrial robotics and automation firms have shown interest in the approach, and Knight argues the speed could soon have a large impact on manufacturing and e-commerce. At the Conference on Robot Learning in Austin this week, Fan used his keynote to show Astra taking control of a robot arm and painting the Golden Gate Bridge with paint and brushes that were lying nearby.

The consequence is a change in where mistakes live. When a robot behaviour costs hours instead of weeks, teams will train far more of them, and the scarce step becomes proving which ones are safe outside the simulator. A treat that misses Leona’s bowl is a cheap way to learn that lesson. A warehouse arm running a model-written skill that nobody has checked on the real floor is an expensive one.

Tags: nvidia, Hugging Face, Microduck, Pollen Robotics, Robots