Viral videos of humanoid robots kicking children or spiraling out of control have exposed a fundamental industry problem: how to deploy a machine in a warehouse or office without injuring a person. Manufacturers are responding to this challenge with multi-layered safety systems—from specialized Nvidia chips to completely abandoning bipedal designs in favor of wheeled platforms.
The industry has encountered a paradox: the smarter and more autonomous a robot becomes, the harder it is to guarantee its safety. Traditional industrial manipulators are deterministic systems operating on a rigid set of rules. Humanoid machines controlled by AI, however, are probabilistic systems: their actions are based on statistical models, not absolute predetermination. It is this uncertainty that creates the main threat.
Multi-layered protection: from chips to infrastructure
Nvidia has already introduced a solution for humanoid robots based on Blackwell chips. As explained by the company's senior director of robotics, Amit Goel, the safety system interprets sensor data, identifies potential hazards, and can instantly stop the machine. The key innovation is the creation of a separate software stack where the functional system and the safety system operate in constant dialogue, rather than in isolation.
Another approach is offered by Fort Robotics: their controllers gather information not only from onboard cameras but also from external infrastructure. The company's CEO, Samuel Reeves, emphasizes that it is now necessary to analyze not just the presence of a person in the work zone, but their posture, trajectory, and the degree of reliability of this data for decision-making.
Standards lag behind—manufacturers act on their own
The problem of stability loss is so critical that the International Organization for Standardization (ISO) has created a separate expert group. However, the publication of unified requirements is not expected until mid-2028. In the meantime, companies are developing their own scenarios.
German company Neura Robotics is releasing the 80-kilogram bipedal robot 4NE1. Its design includes a "controlled fall": if a knee joint fails, the machine attempts to regain balance, and if that is impossible, it folds downward like a collapsing building, minimizing damage.
A more radical solution was chosen by Dexmate: their robots stand on wheeled platforms with a low center of gravity. The battery and electronics are located in the base, eliminating falling altogether. Long manipulators reach items on shelves, but the machine itself remains static.
Energy under control
Brad Porter, founder of Cobot, suggests looking at the threat soberly. His wheeled robots, pushing carts in hospitals or sorting parts in factories, move at walking speed and do not have super-strong grips. "We don't need to put a lot of energy into actions. We're not trying to crush watermelons," he notes.
My analysis: The humanoid robot market is at a crossroads. On one hand, progress in AI opens up incredible opportunities for autonomy. On the other, it is precisely this autonomy that creates non-deterministic risks that frighten businesses and regulators. Until ISO releases unified standards, the leaders will be those companies that can offer not just a smart robot, but a predictably safe one. Abandoning legs in favor of wheels may seem like a step backward, but in a warehouse or hospital setting, it is a pragmatic solution that will save not only reputations but also lives.