Parametric Is Teaching Robots to Fold Your Laundry - and to Run the Whole Business

The pitch fits on a laundry tag: dirty clothes in, folded clean clothes out. What happens in between is where Parametric gets interesting. Inside one of the larger wash-and-fold operations in San Francisco, a wheeled robot with two arms picks up a towel, folds it, and sets it down. When it gets the fold wrong, a customer notices. That complaint does not go into a suggestion box. It goes into a training pipeline, and the robot gets a little better at folding towels.

That loop - work, feedback, improvement - is the whole company. Parametric, part of Y Combinator's Fall 2025 batch, describes itself in five words: physical businesses that run themselves. The founders are not trying to build a general-purpose robot that can do anything. They picked one boring, high-volume job and rebuilt the business around a machine.

It is a narrow beginning, and that is the point. Robotics has a long history of impressive demonstrations that never survived contact with a real workday. A robot that folds a towel perfectly under stage lights is a different animal from one that folds ten thousand towels a week, in every size and fabric a walk-in customer happens to bring. Parametric's answer to that gap is not a cleverer arm. It is a place to practice: a live business, with paying customers, generating the exact stream of feedback the system needs to keep improving.

The idea

Own the business, not the robot

Most robotics companies sell hardware. A factory buys an arm, a warehouse leases a fleet, and the vendor moves on to the next sale. Parametric took a different position. It is structured as a public benefit corporation, and instead of shipping robots to customers, it operates the business the robots make possible - and keeps the operating margin.

The first business is laundry. Not because folding is glamorous, but because it is the opposite: repetitive, measurable, and everywhere. A towel folded correctly looks like a towel folded correctly. There is a clear right answer, which turns out to matter a great deal when you are trying to teach a machine.

The other design choice is subtle. Parametric does not retrofit a robot into a workflow built for human hands. It redesigns the workflow so the robot is a first-class citizen. The hardware - a mobile base on wheels, two arms built for long uptime and high throughput - is shaped around the task, not squeezed into a room that was shaped around people.

How it works

The feedback loop that chatbots got, and robots didn't

Anyone who has used a modern chatbot has benefited from a technique called reinforcement learning from human feedback, or RLHF. A model produces an answer, a human rates it, and that rating nudges the model toward better answers over time. Language labs built enormous machinery around this loop. Robots never really had one.

Parametric built it for the physical world. The company runs an automated pipeline that combines real customer feedback with a judge model - a second model that scores the robot's work - and uses that signal to reinforcement-learning fine-tune the robot's policy. The customer defines what "good" means. The judge translates it into a reward. The robot chases the reward.

🧹

Robot acts

Folds, sorts, handles laundry on-site

💬

Customer feedback

Real customers flag what's off

⚖

Judge model

Scores work into a reward signal

🎯

RL fine-tune

Policy steers toward "good"

The loop repeats - each cycle the robot folds a little better

The practical payoff is speed of learning. Parametric says its robots can pick up a new behavior on-site with less than an hour of data - no human demonstrator patiently guiding the arms through a thousand repetitions. In a field where data is the expensive part, learning fast from a live business is closer to a moat than a bigger model would be.

The receipts

Beating the state of the art on matched hardware

Claims are cheap in robotics, where staged demos are a genre unto themselves. Parametric's are specific enough to check. On matched hardware and benchmarks, the company reports its robots beat Physical Intelligence's Pi0.5 model - a leading vision-language-action model at the time - with roughly 3x higher reliability and 20% greater speed. It says it got there not by scaling up someone else's foundation model, but by building and training its own dense reward models.

Parametric

3x reliable

Pi0.5 baseline

1x

There is a research flourish underneath the folding, too. Parametric says it trained the first transcoders on a vision-language-action model. Transcoders are an interpretability tool - a way of reading what is happening inside a model - and the lineage traces back to the same interpretability work that produced curiosities like "Golden Gate Claude." Parametric's version lets the system predict a robot's intended action before the arm moves. Research toys, in other words, becoming product features.

Who's building it

From recommendations and self-driving to towels

The founders arrived at laundry from two directions that both taught the same lesson. Chief executive Cody Swain led model ROI and offline simulation for recommendations at Meta, then spent time in crypto building a stablecoin wallet and zk-SNARK identity proofs. Chief technology officer John Newsom led the self-driving team at Parallel Systems, a Series B startup, and researched robotics, world models, and large language models at Berkeley's BAIR lab.

Recommendations and autonomous driving look nothing alike until you notice what makes both work: a tight feedback loop that tells the model, over and over, what counts as a good decision. A feed learns from clicks. A self-driving stack learns from interventions and near-misses. Robots doing physical work never had that loop wired up - the feedback was trapped in a customer's head, or a shift manager's grumble, and never made it back to the machine. Parametric's founding bet is that closing that gap is the unlock.

The choice to run a laundromat rather than license a platform also keeps the team honest. When you operate the business, a bad fold is your problem, not a slide in someone else's deck. The team - a small, research-heavy group in San Francisco - lives with the output of its own models every day, which is a blunt but effective way to keep the benchmarks tethered to reality.

The landscape

Where Parametric sits

The physical-AI field is crowded with well-funded ambition. Companies like Physical Intelligence chase general-purpose robot foundation models; humanoid outfits raise on the promise of a machine that can do everything. Parametric is deliberately narrower. It is not selling a brain or a body to other people. It is running a specific operation and improving one task at a time from real customer feedback.

That focus is the differentiator. A general model has to be adequate at everything. Parametric only has to be excellent at the handful of things its laundromat needs, and it gets to practice on live orders every day. Against traditional fixed automation - the conveyor-and-clamp machinery of an industrial laundry - the pitch is flexibility: a system that learns new items and new folds without a re-engineering project.

What you can do with it

If it works, the model travels

For now, the thing a customer can actually do with Parametric is drop off laundry and get it back folded, in a San Francisco shop where the folding is increasingly done by a robot that is learning on the job. The larger idea is portable. Any labor-intensive, repetitive operation with a clear notion of a job done right - the kind of work that is easy to describe and tedious to do - is a candidate to be redesigned around a machine that learns from the people it serves.

Whether the economics hold at scale is the open question, and Parametric is early enough that its most impressive numbers are still its own. But the shape of the bet is unusually clear for a robotics startup: pick a boring business, wire up the feedback loop that everyone else skipped, and let the customers do the teaching.