Humanoid robots are shipping at scale, but safety validation isn't keeping up

Somewhere between 19,000 and 22,000 humanoid robots shipped globally in the first half of 2026, depending on whose count you trust. Most of them came from China. AgiBot and Unitree alone account for roughly three-quarters of the market, with Galbot, UBTech and Leju rounding out the top five. AgiBot says its 15,000th robot rolled off the line in June. Unitree went public on Shanghai's STAR Market in August.

Those robots are no longer just dancing on stage. AgiBot G2s have been livestreamed working a real Longcheer production line, UBTech's Walker S has trained in auto plants for NIO, Zeekr and BYD, and Xiaomi put a humanoid "intern" in its own car factory.

In the U.S., the story right now is robots getting out of the cage. 

Agility's Digit 5 is designed to work beside people with no fencing, and Boston Dynamics pitches "fenceless guarding" for its production Atlas. And 1X plans to ship 20,000 Neo robots into homes starting this year, where there are no safety zones at all.

The rules are lagging on every continent.

China moved first on paper: in February, MIIT's new national technical committee released the country's first standards system for humanoid robots and embodied AI, with safety and ethics as one of six pillars. But a standards framework is a map of standards to come, not a test a robot can pass today. The international equivalent, ISO 25785-1, is still a committee draft. In Europe, London-based Humanoid is betting on wheels in part because wheeled robots can be certified under existing rules while bipedal ones can't yet. And local governments are starting to improvise: San Mateo County in California has approved rules requiring permits and kill switches for commercial humanoids.

Add the U.S. FCC ban on Chinese robots, and safety and trust are quickly becoming as much a competitive weapon as payload or price.

So how do you prove a humanoid is safe before it's standing next to a person?

I put that question and several others to Erikk Cass, AVP of autonomous mobility at iMerit, which provides data annotation and model evaluation services to eight of the world's leading autonomous vehicle companies and a growing list of robotics firms. His core argument: humanoids need the same failure-to-training-data feedback loop the AV industry spent 15 years building, and they're about a decade behind.

Here's the full email interview, unedited.

John Koetsier: What are the hardest safety edge cases for humanoid robots?

Erikk Cass: The hardest cases are the ones where several uncertainties stack up at once: a person moves in a way the robot didn't predict, an object is partly hidden, the lighting or layout changes, and the robot is looking at something it never saw in training. Any one of those is manageable. Together they push the model into territory where its confidence should drop, and the critical question is whether it recognizes that and fails safely. A common misconception is that the answer is a smarter model that can account for everything. In practice, safety comes from how quickly a system learns from the situations it got wrong, which means every near miss has to become labeled training and validation data, not just an incident report.

John Koetsier: Autonomous vehicles have billions of miles of simulation, road data. What's the equivalent for humanoids?

Erikk Cass: Humanoids need the same feedback loop the AV industry spent fifteen years and billions of dollars building: simulation, real sensor data, and a disciplined process for turning real-world failures into training and validation data. The difference is scope. An autonomous vehicle mostly needs to understand where things are and where they're going. A humanoid also has to understand manipulation, human behavior, and the consequences of getting a task wrong, in thousands of environments rather than a road network. So the data problem is larger and more multimodal, and the industry is roughly where AVs were a decade ago in terms of how much of that data exists. There is no shortcut through it.

John Koetsier: We're working on robot safety in warehouses and factories ... what about the home?

Erikk Cass: A factory is a controlled environment. You define the workflows, the operating zones, and who is allowed near the robot, and the job is largely the same every day. A home has none of that. It has children, pets, clutter, stairs, and people who behave unpredictably, and the layout changes every day. The robot has to recognize far more uncertainty and, just as importantly, know when not to act. That "do nothing safely" behavior is much harder to design and validate than it sounds, and it is where I'd expect the first serious problems to show up.

John Koetsier: Is collision avoidance enough, or does a safe robot need to understand human intent?

Erikk Cass: Collision avoidance is necessary, but it is one layer, and it's the last one. If a robot is relying on collision avoidance, something upstream has already failed. A robot working around people needs enough contextual understanding to anticipate what a person is likely to do next, recognize when a situation is ambiguous, and choose a safe fallback when it isn't sure, rather than reacting in the final half second.

John Koetsier: How do you test an AI model whose behavior can change as the software is updated or retrained?

Erikk Cass: You stop thinking of testing as a gate and start thinking of it as a loop. Every meaningful update gets run against a regression set of known failures, edge cases, and realistic simulations before it ships, and every new failure in deployment goes back into that set. This is also where human expertise stays essential. Automated metrics can tell you a model's performance changed. They can't tell you why the robot hesitated in a doorway or misjudged a hand reaching toward it. Someone with domain knowledge has to look at that scenario, decide what the correct behavior was, and label it so the next version learns it. That judgment is not something you can retrain your way out of.

John Koetsier: Should there be standardized safety benchmarks for physical AI, similar to crash tests for cars?

Erikk Cass: Common baseline benchmarks would help, particularly for failure behavior, how a system handles uncertainty, and human interaction. But crash tests work because every car faces roughly the same physics. Physical AI doesn't have that. A warehouse robot, a home robot, and a surgical robot operate in different environments with very different consequences when something goes wrong, so any standard will need a common core plus domain-specific tests. The AV industry's experience with scenario-based validation is a reasonable starting point.

John Koetsier: Are companies moving humanoids into human environments faster than the safety infrastructure can keep up?

Erikk Cass: Yes, the technology is moving faster than the evaluation infrastructure. A polished demo tells you what a robot does under expected conditions. It tells you nothing about the rare, ambiguous situations, and those are the ones that determine whether the robot is actually safe around people. The question I'd ask any company shipping humanoids is not "can you show me it working" but "can you show me how it fails, and what you did with that data afterward."

John Koetsier: What will be our biggest safety problem in humanoid robotics?

Erikk Cass: The long tail of situations nobody anticipated. A model can handle the vast majority of cases very well, and public trust will still be decided by what happens in the unusual ones, because those are the incidents that get filmed. The AV industry learned this the hard way. One well-publicized failure can set a program back years regardless of how good the aggregate numbers are, which is why real-world stress testing has to happen before deployment, not after.

John Koetsier: What evidence would you need that it is safe?

Erikk Cass: I'd want to see performance that is consistent across normal and adverse conditions, not just a highlight reel. Specifically, I'd want evidence of how the robot behaves when sensors degrade, when people act unpredictably, and when it encounters something outside its training data, along with a documented record showing those failures went back into training and validation. That closed loop is the evidence. A company that can show it is one I'd trust around people. A company that can't is asking you to take the demo on faith.

John Koetsier: Thank you for your time!