Humanoid Robots Struggle Outside Controlled Environments
Humanoid robots are making headlines, but often for the wrong reasons. Recently, a robot at a San Jose restaurant went haywire, dancing uncontrollably, smashing plates, and sending tableware flying. Staff had to physically restrain the machine, struggling to find a way to shut it down.
This incident wasn’t isolated. In Macau, police escorted a humanoid robot off the street after it frightened an elderly woman, leading to her hospitalization. The robot, valued at around $13,500, had been following her silently. This event was quickly dubbed the “world’s first robot arrest.”
The Gap Between Demos and Deployment
Billions of dollars are being invested in humanoid robots, with China alone boasting over 140 manufacturers and 330 models. These robots can perform impressive feats like backflips and kung fu, often showcased on national television. However, their performance plummets when placed in real-world settings with unpredictable people and environments.
The core issue is that these robots are built for demonstrations, not for actual deployment. The difference between a controlled stage and a busy restaurant is a significant engineering challenge that current investments haven’t fully overcome. Robots often fail because the real world doesn’t follow a script.
Recent Failures Highlight Engineering Challenges
A recent example is Xpeng Motors’ humanoid robot, Iron, which fell on its face during its public debut. The company’s CEO compared the fall to a child learning to walk, but this was the robot’s first public appearance. This reflects a trend where Chinese robotics companies are testing prototypes in public, with less concern for embarrassment and more focus on rapid iteration.
The Unitree H1 robot’s chaotic stage performance was attributed to a tether attached to its head, which confused its balance algorithms. This led to a feedback loop of increasingly large corrections, causing it to flail violently. Engineers apparently had not tested the robot with such a basic physical constraint.
At a Beijing half marathon, 21 humanoid robots participated, but most failed to finish. Some collapsed early, others overheated, and one crashed into a wall. Only six completed the race, with the fastest requiring multiple battery swaps and still falling once.
Why Walking is So Hard for Robots
These incidents reveal a critical truth: the conditions that make a robot look impressive are rarely the same as those where it is genuinely useful. Walking on two legs is incredibly energy-intensive for a machine. Unlike wheeled robots, bipedal systems require constant, tiny adjustments just to stay balanced.
Dozens of motors work every second to prevent a humanoid robot from toppling over, which quickly drains its batteries. This explains why the dancing robot incident was so revealing. It wasn’t necessarily a bad AI problem, but a lack of a hardware kill switch. Staff had to physically restrain it while trying to access an app to shut it down, highlighting a basic safety feature’s absence.
The Scale of Chinese Investment
Despite these challenges, China is pouring massive resources into humanoid robots. In 2025 alone, Chinese investors injected approximately $5.5 billion into embodied intelligence startups, a 326% increase year-over-year. Unitree Robotics, the company behind the problematic G1 and H1 models, was valued at $7 billion during its IPO preparations.
China’s state council predicts the humanoid robot market will exceed one trillion yuan by 2035. Chinese firms already dominate the global market, with AI Bot shipping over 5,000 units in 2025. However, it remains unclear how many of these are actual commercial sales versus demo units or pilot programs.
A Strategy of Speed Over Safety
China’s national strategy prioritizes AI, with embodied AI listed as a top national priority. The approach is to build and ship quickly, addressing problems later. A report by Bane and Company noted that current demonstrations often hide technical limitations by using controlled environments or remote oversight.
The gap between what is shown at research conferences and what can be deployed in real facilities is widening. Industry insiders admit that for every $100 spent on a robotic solution, only about $20 goes to the robot itself. The remaining $80 covers safety infrastructure, integration, and environmental retrofitting.
Why This Matters: Safety Concerns in Public Spaces
The Macau incident, where a Unitree G1 robot followed an elderly woman, was not an autonomous failure but a result of remote control. The robot couldn’t navigate around her, so it just stood there, causing her to be hospitalized for shock. This robot was eventually returned to its operator with a warning.
These robots are already appearing in public spaces across China. In Shenzhen, a robot named T800 patrols tourist areas with police. In Shanghai, a robot named Xiao helps direct traffic. Crucially, there are no clear safety standards governing these deployments.
A safety expert noted at CES 2026 that while some robots use physical separation or move very slowly, those that move at human speed and interact directly with people have significant safety gaps that need urgent attention. The International Organization for Standardization is still developing ISO 257785, the first standard for humanoid robot safety, which is not yet complete.
The Future of Humanoid Robots
China is building humanoid robots rapidly and at scale, similar to its electric vehicle industry. With over 140 companies and a trillion-yuan market goal, the ambition is clear. However, incidents like dancing robots that can’t be turned off, robots frightening elderly citizens, and robots failing at their own launch events are not just amusing internet content.
They represent the consequence of treating demonstrations as finished products. The real world is unpredictable, lacking controlled stages, flat surfaces, and engineers ready with laptops. It includes dark streets, people on their phones, and messy restaurant tables. The investment and ambition are substantial, but until robots can reliably navigate reality and include basic safety features like an off switch, they will continue to stumble.
Source: The Real Reason Humanoid Robots Keep Failing in the Real World (YouTube)
