Our Take
Physical AI companies are treating cities as open-air labs—and regulators haven't caught up to the speed of deployment.
Why it matters
Robot startups can no longer rely on simulation to train systems that must navigate unpredictable human environments. This acceleration toward real-world testing forces a reckoning: either the sector self-governs deployment practices, or cities will impose friction after incidents occur.
Do this week
Robotics teams: document your testing protocols and incident logs now, before local authorities ask for them—compliance will follow deployment, not precede it.
Startups are deploying robots in live cities to train physical AI
A wave of physical AI companies are moving beyond closed labs and controlled facilities, instead using city streets, retail spaces, and public environments as active training grounds. The AP reporting captures entrepreneurs treating real-world deployments as the primary path to building systems that can handle unpredictable human behavior, infrastructure variation, and edge cases that simulation cannot replicate.
This is not a theoretical shift. Humanoid robots, autonomous delivery systems, and vision-based agents are now operating in uncontrolled settings where their failures are visible and their data collection happens alongside foot traffic, weather variation, and human interference. Companies are learning that the gap between simulated environments and real physics is too wide to bridge with software alone.
The move reflects a hard truth in robotics: you cannot train a physically capable system without physical data. Simulation has hit its limits for systems that must grasp objects, navigate crowded spaces, or respond to unpredictable human interaction. Real-world deployment becomes the only way to collect the signal needed for the next generation of models.
Regulatory frameworks are nowhere near this pace
Cities and states have not moved to regulate robot deployment in public spaces at the speed startups are moving to test them. Local authorities lack clear standards for what constitutes safe testing, incident reporting, or liability. The AP framing hints at the vacuum: entrepreneurs are staging systems in high-traffic areas without formal approval pathways.
This is not necessarily reckless. It is typical of infrastructure transitions where private capital moves faster than public governance. Electric scooters faced this same pattern. The risk is that a high-profile failure (a robot injuring a pedestrian, blocking emergency response, or causing property damage) will trigger reactive, blunt regulation that stifles the sector rather than focusing it.
Companies deploying in this window have a window to establish good practices voluntarily. Those that document safety protocols, report incidents transparently, and demonstrate responsible scaling will have credibility when regulators arrive. Those that treat cities as wild west testing grounds will be used as the cautionary example in the first lawsuit.
What builders should do now
If you are operating robots in public spaces, your liability exposure is real even if your legal framework is unclear. Start now: log every failure mode, every human interaction that diverged from your expected behavior, every environmental condition that caused deviation. This data serves dual purposes: it improves your models, and it becomes your defense in any later dispute over whether you were operating responsibly.
Second, engage with local government before you need to. A conversation with your city's transportation or public safety office about what you are doing, how you mitigate risk, and what you report is not a constraint on your timeline. It is insurance against surprise enforcement later.
Third, separate training deployments from commercial deployments in your communication with the public. Be transparent about what is a test versus what is a production system. This distinction will matter enormously when liability questions arise.
The companies that succeed in this transition will be those that treat real-world deployment as engineering work, not just a staging ground. That means discipline around safety, transparency about incidents, and proactive engagement with the communities where you operate.