Viral videos of humanoid robots kicking children or dancing uncontrollably have exposed a fundamental industry problem: how to deploy a machine on a production floor without injuring people. Incidents involving severe injuries and even fatalities have already been recorded, forcing developers to radically rethink safety approaches while simultaneously slowing down large-scale adoption.
The key challenge lies in the architecture of modern machines. Traditional industrial robots—welding machines or forklifts—are deterministic systems operating on a rigid set of rules. Humanoids, driven by AI, are probabilistic systems. They make decisions based on statistical probabilities, making their behavior less predictable and requiring fundamentally different safety mechanisms.
Multi-layered Protection: From Chips to Infrastructure
Nvidia has already introduced a specialized safety system for humanoid robots based on Blackwell chips. As explained by the company's Senior Director of Robotics, Amit Goel, the new model can interpret sensor data about potential threats and instantly stop the robot in unsafe conditions. This is not just a software patch but a fundamental layer of the operating system that ensures constant interaction between functional and safety protocols.
Another approach is offered by Fort Robotics from Philadelphia. They are developing controllers that gather information from multiple sources—not only from the robot's own cameras but also from the surrounding infrastructure. CEO Samuel Reeves emphasizes that it's not just about detecting a person in the work zone, but about complex analysis: where they are, in what posture, and how much this data can be trusted for decision-making.
Standards and Engineering Solutions
The problem of stability loss is so critical that the International Organization for Standardization (ISO) has created a separate expert group to study this issue. Unified requirements are expected to emerge no earlier than mid-2028. In the meantime, manufacturers are seeking their own solutions.
German company Neura Robotics, which produces the 80-kilogram bipedal robot 4NE1, has implemented a "controlled fall" algorithm. If the system diagnoses a failure, for example, in a knee joint, the robot first tries to regain balance, and upon failure, it folds downward like a collapsing building, minimizing damage.
Others have gone even further, eliminating the source of risk altogether. The startup Dexmate creates robots on wheeled platforms with long manipulators. The battery and electronics are placed in the base, providing a low center of gravity and preventing falls. Cobot founder Brad Porter suggests looking at the problem without excessive drama: his wheeled robots, pushing carts in hospitals, move at walking speed and do not have an ultra-strong grip. "We don't need to put a lot of energy into every action. We're not trying to crush watermelons," he jokes.
Cryptalist Analysis: The current situation resembles the early days of autonomous vehicles, when every failure made headlines. For humanoid robotics, the "valley of death" between experimental prototypes and industrial safety will only be overcome through a combination of hardware constraints (such as abandoning legs) and software safety nets. Nvidia's investments in this segment are a clear signal to the market: safety will become either the main driver or the main barrier to mass adoption.