Safety of Humanoid Robots: From Viral Falls to Multi-Layered Protection
Viral videos of robots kicking children or dancing uncontrollably have exposed a fundamental problem: how to deploy a machine in a warehouse without injuring a person. The industry is responding to this challenge by creating multi-layered safety systems—from specialized Nvidia chips to abandoning bipedal designs in favor of wheeled platforms.
These incidents are forcing developers to rethink safety mechanisms, which inevitably slows down the mass adoption of humanoid robots. Cases of severe injuries and even fatalities during interactions with such devices have already been recorded. As noted by Michelle Silva, a functional safety expert at Reynolds & Moore, even simply cutting power can be dangerous: if a humanoid robot loses power, it could collapse and crush a person.
The key difficulty lies in the architecture. Traditional industrial robots are deterministic systems that strictly follow a given set of rules. Humanoid robots, however, use AI and are probabilistic: their actions are based on statistical probabilities rather than rigid certainty. This is why they require a fundamentally different, multi-layered approach to safety.
Nvidia has already proposed a solution—a safety system based on Blackwell chips. As explained by the company's Senior Director of Robotics, Amit Goel, the model analyzes sensor data on potential threats and can instantly stop the robot in unsafe conditions. He emphasized that the safety system and the functional system must constantly interact within a broad context, and Nvidia has created a software stack that allows them to work together.
Another layer of protection involves the use of external infrastructure. Fort Robotics develops controllers that gather information from multiple sources. CEO Samuel Reeves explained that it is no longer just about detecting a person's presence, but about complex analysis: where they are, in what posture, and how much that data can be trusted for decision-making.
The problem of stability loss is so pressing that the International Organization for Standardization has created a separate expert group. The publication of unified requirements is not expected until mid-2028, and in the meantime, manufacturers are developing their own scenarios. German company Neura Robotics, which produces the 80 kg bipedal robot 4NE1, has implemented a "collapsing building" mechanism: in the event of a failure, for example, in a knee joint, the machine tries to regain balance, and if that is impossible, it falls by folding downwards, minimizing damage.
Some developers have taken a radical path. Dexmate creates robots on wheeled platforms with long manipulators. Co-founder Yuzhe Qin noted that the battery and electronics are placed in the platform, ensuring a low center of gravity and completely eliminating falls. Cobot founder Brad Porter, whose wheeled robots push carts in hospitals, calls for a sober assessment of the threat: his machines move at walking speed and do not have super-strong grips, as their tasks do not require "crushing watermelons."
Expert opinion: While the industry moves toward standardization, we are witnessing a classic "period of throwing stones." Abandoning legs in favor of wheels is not a technological regression, but a pragmatic step toward commercialization. However, it is the "probabilistic" nature of AI-powered humanoid robots that will become the main headache for regulators. Entrusting a machine that "thinks" statistically with a human life is a task that requires not just new chips, but a complete overhaul of the philosophy of industrial safety.