A San Francisco startup briefly advertised remote shifts paying $20 an hour to steer robot dogs through the city’s streets. The experiment aimed to capture real-world sensor data for AI training, revealing that algorithms still struggle to handle physical hardware without human intervention.
The job posting lasted only briefly before disappearing from the internet, but it pulled back the curtain on how tech firms are tackling the messy realities of robot autonomy. Dimensional, a startup based in San Francisco’s Dogpatch neighborhood, posted an online advertisement seeking workers willing to take four-legged machines on multi-hour walks across city streets, as reported by Hoodline.
The listing drew widespread attention for its unusual premise: paying humans to walk mechanical pets.
Yet the mechanics behind the test point to a persistent bottleneck in robotics research. While artificial intelligence can write code and generate text with ease, guiding physical machinery through unpredictable urban environments remains a severe challenge for automated systems.
Inside the Dogpatch Robot Walks and the Unitree Machines
The deleted listing asked for dependable applicants to pilot four-legged units on outdoor testing routes lasting four to six hours at a stretch. The company required no prior robotics experience, offering paid training to get operators up to speed quickly.
The work involved steering a roughly 33-pound machine along designated paths in San Francisco while onboard cameras and sensors logged surroundings, street conditions, and traffic. The hardware itself consists of commercial robot dogs manufactured by Chinese robotics firm Unitree. These units retail for about $2,000 each and accept commands from a handheld remote or a mobile application, according to Fox Business reporting.
Walking the machines was far from a casual stroll. Because the hardware occasionally overheated or malfunctioned mid-route, human operators had to be physically capable of carrying the 33-pound robots back to the Dogpatch office by hand.
Stash Pomichter Gathers Ground Truth Data
The street tests were not designed to exercise mechanical pets. Dimensional founder Stash Pomichter explained that steering the units manually generates what the company terms “ground truth” data—the raw sensor readings, video feeds, and spatial measurements necessary to train software for future autonomous driving.
To measure progress, the startup compared how effectively the robots performed under human control against trials driven entirely by AI agents. The results were stark.
That conclusion mirrors observations from similar tests conducted by Anthropic, which found that language models struggled significantly when attempting to manipulate robot joints directly, showing better performance only when directing pre-trained low-level controllers.
Dimensional Develops Open-Source OS for Diverse Robot Fleets
Dimensional’s long-term ambitions extend well beyond experimental walks. The company is actively developing dimOS, an open-source operating system currently available as a prerelease beta. The platform aims to let AI agents control a diverse array of hardware—including quadrupeds, humanoid robots, and robotic arms—using natural language commands.
To build out its fleet, the startup has purchased more than 100 robot dogs and 50 humanoid machines from manufacturers across the United States, China, and Germany.
The ultimate goal is deploying these automated workers into demanding environments such as factories, data centers, construction sites, and security patrols. A hint of that future is already taking shape: a robot dog currently patrols an Atlanta apartment complex under the remote supervision of an operator based in the Philippines.
Yet deploying autonomous machinery onto public sidewalks brings immediate friction. The street tests sparked concerns over pedestrian safety, data privacy, and the continuous recording of public spaces by private sensors. Hostile messages directed at the company over potential misuse became severe enough that Dimensional referred them to local police. As startups race to build autonomous physical systems, they continue to run up against basic physical limits—including battery life, thermal shutdowns, and the unanswered regulatory questions of who holds liability when a robot causes harm in public.