The $20 an Hour Job Teaching Robot Dogs How to Walk a City

A gray Unitree Go2 robot stands on snow during a mountain-rescue technology demonstration.
A Unitree Go2 during a Salvamont demonstration in Romania in March 2025. Frame from a HotNews Romania video; illustrative hardware, photographed away from the San Francisco project. Image credit: HotNews Romania — Adi Iacob and Ovidiu Popica, CC BY 3.0. Resized, full frame retained.

A robot dog that needs a person to take it for a walk sounds like a joke about the future arriving slightly out of order. I find it a useful picture of where physical AI actually stands. A machine can look impressively capable in a short demonstration while still depending on somebody to steer it, recover it and explain what went wrong outside the demonstration.

Dimensional, a San Francisco robotics startup, advertised robot-walking work at $20 an hour, approximately €17.83. The advertisement described four to six hours of walking and required applicants to handle equipment weighing about 33 pounds, or 15 kilograms. Kashmir Hill’s reporting for The New York Times, republished by The Star, says the advertisement was subsequently deleted. This is a reported experiment in collecting robot training data, rather than a verified vacancy readers can apply for today. Kashmir Hill’s reporting for The New York Times · ECB reference rate, November 10

The interesting part is what the person contributes. Walking a familiar street is full of small judgments that we rarely bother to describe. Teaching a machine to navigate that street turns those judgments into work somebody has to do, record and evaluate.

What the walker is actually teaching

According to Hill’s account, Dimensional’s walkers guide the machines while cameras and sensors record their journeys. The reporting also describes overheating that can leave a walker carrying the robot back. Those details make the job more concrete: the person is an operator and a recovery plan, as well as a source of examples. Kashmir Hill’s reporting for The New York Times

The company’s own dimOS documentation helps explain the arrangement. Its software supports Unitree Go2 robots, remote operation and autonomous navigation components. The remote-control interface includes a camera view, a map and manual driving controls. It also permits a person to interrupt a navigation plan. These are documented capabilities of the software, not evidence that every sidewalk journey has become autonomous. Dimensional’s official documentation · Dimensional’s remote-operation guide

Imagine a delivery van temporarily blocking the pavement. The useful lesson is more than “move left.” A walker has to notice whether the alternative route enters traffic, whether someone with a pushchair needs the remaining space, and whether waiting is the better decision. Even describing that episode properly takes context. A video of the successful detour might omit the hesitation that made it safe.

For me, this is the difference between a memorable demonstration and a dependable service. I want to know what happens on the ordinary bad afternoon: the pavement is crowded, a sensor gets confused, the battery is running down and the obvious route has disappeared. Those moments are part of the product.

Why walking around is valuable data

Dimensional documents tools that store sensor observations with timestamps and replay them later. That matters because developers can revisit a recorded situation while changing the software. A recording can help them inspect what the robot saw, when it saw it and how a different component processes the same observations. Sensor recording and replay

Replay does not recreate the entire physical world. If new software chooses a different turn, the original recording cannot reveal everything that would have happened along that untraveled route. I see recordings as useful evidence for development, with fresh field work still needed to establish how the changed system behaves outside.

A worthwhile collection plan also needs variety. Ten long walks along an easy route may tell an engineer less about a particular weakness than one carefully documented difficult junction. Conversely, collecting only dramatic failures could produce a misleading picture of everyday reliability. The dataset needs the uneventful stretches that establish what normal operation looks like.

I would therefore judge this kind of project by the questions its walks can answer. Did the operator intervene? Why? Was the problem perception, route selection, physical footing or a hardware interruption? If every incident is simply labeled “human took over,” a great deal of useful information disappears. Distance covered is an appealing progress figure, but it is only one measurement.

The hourly rate deserves ordinary arithmetic

The novelty of the job makes $20 an hour sound like payment for an amusing afternoon. I prefer to calculate it as work. At the advertised rate, four paid hours would produce $80, about €71.34; six would produce $120, about €107.01. These are gross arithmetic examples, before any taxes or expenses. The euro equivalents use the November 10 ECB reference of $1.1214 per euro, so they are conversions rather than European wage offers. ECB reference rate, November 10

A hypothetical one-hour round-trip commute changes the four-hour example to five hours committed to earning that $80. That is $16, around €14.27, per hour of the worker’s total time before expenses. This does not establish whether the advertised job was attractive in San Francisco. It shows why a memorable headline rate does not settle the question.

The terms I would want in writing are quite practical: which activities count as paid time, how breaks and equipment interruptions are handled, who provides training, and what happens when a machine needs recovery. Carrying a 15-kilogram object back to an office is a different commitment from carrying a phone and a controller.

Questions I would resolve before accepting similar work
Part of the jobWhat the written terms should clarify
Paid timeWalking, setup, charging delays and incident notes.
RecoveryTraining and support when a machine cannot continue.
TravelWhere the shift starts and which trips are covered.
Data qualityHow careful interventions and useful notes are assessed.

Those questions also protect the quality of the data. If workers are rewarded only for distance, stopping to document a confusing situation could look unproductive. If the company values careful intervention notes, the incentives change. I have not seen Dimensional’s employment terms or its internal payment structure, so this is how I would assess such an offer, rather than a claim about its policies.

The price of the robot is only the beginning

A hardware purchase is easy to photograph and count. The supporting work is less visible: charging, repairs, software setup, supervision, storing recordings and understanding failures. An inexpensive robot can become an expensive service if it frequently needs a person to rescue or reset it.

Consider a hypothetical business that wants a robot to carry supplies between two buildings. The right comparison includes how often the route is completed, how much human supervision it needs and what happens when the route changes. A machine that succeeds on most trips but requires an unpredictable recovery can create a scheduling problem even when its average performance looks respectable.

The economics also change if one operator can reliably supervise several machines. That could make the service more affordable. But I would want evidence from the intended environment before assuming that supervision can be stretched. A ratio that works on a quiet private site may be inappropriate on a crowded public pavement.

This is why I am interested in the operating system as much as the robot. A common software layer could make it easier to connect hardware, record incidents and replace individual components. Dimensional publishes its project and documentation openly, which makes the approach more inspectable. Open source by itself does not establish reliability; it gives outsiders something concrete to examine. The dimOS source project

A street is shared space

As a traveler, I tend to notice the things that make a pavement easy or difficult to use: luggage, narrow sections, uneven surfaces and people who need more room to pass. A robot joining that space needs to be judged from the perspective of the people already there.

My acceptance criteria would include predictable stopping, clear human supervision during trials and a way for the operator to move a stopped machine out of the way. I also want the people collecting street footage to explain what they retain and how they handle incidental images of passersby. These are questions for the deployment team; the existence of recording tools does not answer them.

There is an appealing future in a machine taking on an awkward carrying task. I can also see a present in which the machine creates a new carrying task for its handler. Both can be true during development. The useful evidence is whether the second becomes less frequent as the first becomes more dependable.

That is what makes the robot-walking story worth following. It reveals a stage of automation that launch videos often compress into a few seconds: people teaching, supervising and repairing the system. Before I pay for a robot service, I want to know how much of that work is still required, who performs it and whether the quoted price actually covers it.

Comments

Popular posts from this blog

Obsidian’s New Search: Less Friction for Notes in Two Languages

OPPO Find X9 Ultra vs Find X10 Pro Max: Is the Older Ultra Still Worth Buying?

OPPO Find X10 Technical Preview: Cameras, Design and Missing Specs