Viral videos of humanoid robots kicking children or spiraling out of control in warehouses have exposed a fundamental industry problem: how to deploy a machine in the real world without risking human life and health. The answer lies in a multi-layered safety architecture—from specialized Nvidia chips to abandoning bipedal designs in favor of wheeled platforms.
The problem is compounded by a fundamental difference between traditional industrial robots (welding machines, forklifts) and new humanoid systems. The former are deterministic, rigidly following a set algorithm. The latter are probabilistic, operating on AI and statistical models. It is this uncertainty that makes them potentially dangerous: in the event of a power failure, an 80-kilogram machine could simply collapse onto a person.
Technological Solutions: From Chips to Infrastructure
Nvidia has already introduced a safety system for humanoid robots based on Blackwell chips. As explained by the company's Senior Director of Robotics, Amit Goel, the model interprets sensor data about potential threats and can instantly stop the machine. The key innovation is the creation of a separate software stack that allows the functional safety system and the robot's operating system to run in parallel, rather than competing for resources.
Another approach is offered by Fort Robotics from Philadelphia. Their controllers collect data not only from the robot's own cameras and sensors but also from external infrastructure. CEO Samuel Reeves emphasizes that it is now important not just to detect a person in the work zone, but to assess their posture, trajectory, and the degree of reliability of this information. This requires a completely different level of computing power and algorithms.
Abandoning Legs as a Safety Strategy
Some developers have taken a radical path. The company Dexmate creates robots on wheeled platforms with long manipulators. Co-founder Yuzhe Qin explains that the battery and electronics are placed in the base, providing a low center of gravity and eliminating the risk of falling. Such a robot cannot "fall and crush," and its kinematics limit the radius of dangerous impact.
Cobot, for its part, makes wheeled machines that move at walking speed and do not have an ultra-strong grip. "We're not trying to crush watermelons," quips company founder Brad Porter, emphasizing that excessive power is unnecessary for most warehouse and logistics tasks.
Standards Are Lagging Behind
The problem of the lack of unified safety standards is particularly acute. An expert group of the International Organization for Standardization (ISO) is studying the issue of stability loss in humanoid robots, but the publication of requirements is not expected until mid-2028. For now, manufacturers are forced to develop their own scenarios. German company Neura Robotics, for example, has built a "controlled fall" mechanism into its 80-kilogram robot, 4NE1: when a joint fails, the machine folds downwards like a collapsing building, minimizing damage.
Cryptalist Analytics
The humanoid robot market is at a crossroads. On one hand, progress in AI and sensor technology opens up incredible opportunities for automation. On the other, every accident could set the industry back years, triggering regulatory bans and a loss of trust. I believe the key trend for 2025-2026 will not be a race for anthropomorphism, but the creation of hybrid systems where wheeled platforms and manipulators complement each other, rather than trying to copy humans. Safety is not an option, but a foundation without which mass adoption is impossible.