Videos of robots kicking children or spinning out of control while dancing have become more than just viral content — they are a warning signal for the entire industry. The main question facing developers today is: how to deploy a humanoid machine in a warehouse or office without risking human health? The safety problem is becoming the main brake on the mass adoption of humanoid robots, and the industry is seeking solutions on several fronts.
From determinism to probability: a new paradigm of risks
Traditional industrial robots — welders or loaders — are deterministic systems operating according to strict algorithms. Humanoid robots, controlled by artificial intelligence, belong to probabilistic systems. They make decisions based on statistical models, which introduces an element of unpredictability. This is why working in close proximity to humans requires fundamentally different, multi-layered protection.
As functional safety expert Michelle Silva rightly notes, the basic threat lies in physics: if a humanoid robot suddenly loses power, it can simply collapse and crush a nearby person. The weight of such machines often exceeds 80 kg.
Technological barriers: from Nvidia chips to "folding" falls
A key player in creating a safe environment is Nvidia. The company has introduced a specialized safety system for humanoid robots based on Blackwell chips. Senior Director of Robotics Amit Goel explains that the new model interprets sensor data in real time, identifying potential hazards, and can instantly stop the machine. This is not just a software layer, but a full-fledged operating system that allows the functional safety system and the AI control system to work in tandem.
Another approach, being developed by Fort Robotics, involves creating an "intelligent infrastructure." The robot stops relying solely on its own cameras. Instead, controllers collect data from multiple sources, analyzing not just the fact of a person's presence, but their posture, trajectory, and the degree of reliability of this information.
Manufacturers are also seeking mechanical solutions. German company Neura Robotics, which produces the 80-kilogram 4NE1 robot, has implemented a "controlled collapse" algorithm. If the system diagnoses a critical failure, such as a knee joint malfunction, the robot does not fall flat but folds up like a collapsing building, minimizing the impact zone.
A radical solution: abandoning legs
Some companies have gone even further, eliminating the very cause of instability. 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 virtually eliminating falls. This approach is ideal for warehouse tasks where reaching top shelves without the risk of tipping over is required.
Brad Porter, founder of Cobot, urges a rational assessment of the threat. His wheeled robots, which push carts in hospitals, move at pedestrian speed and do not have excessive gripping force. "We're not trying to crush watermelons," he jokes, emphasizing that power-to-weight ratio must strictly match the task.
Analyst's opinion: While the International Organization for Standardization prepares unified requirements for robot stability (expected by mid-2028), the market is moving by trial and error. However, it is precisely now, at the stage of standard formation, that companies have a unique chance to embed safety into the "hardware" and software, rather than trying to "bolt it on" later. The one who solves this problem first will gain not just a technological advantage, but the key to trust from business and society.