Viral videos of robots kicking children and dancing uncontrollably have highlighted a key concern for manufacturers: how to deploy machines in warehouses without endangering people. The industry is addressing these risks with multi-layered safety systems, ranging from Nvidia chips to replacing legs with wheels, according to WSJ.
Various incidents involving robots are prompting developers to enhance safety mechanisms, which in turn slows down mass adoption. Some companies have already reported severe injuries and fatalities involving humanoid devices.
In an interview, Michelle Silva from Reynolds & Moore, a firm specializing in robotic functional safety design, described a basic threat:
“If a humanoid robot loses power, it could fall and crush you.”
Additional complexity arises from AI-based machine architectures. Traditional industrial robots—such as welders, palletizers, and loaders—are deterministic systems that follow a specific set of rules.
Humanoid robots, however, use AI and are probabilistic systems. They operate based on statistical probabilities rather than strict determinism, necessitating multi-layered protection when working alongside humans.
In line with this idea, Nvidia has introduced a safety system for humanoid robots using Blackwell chips. Amit Goel, the company’s senior director for robotics and edge AI ecosystem, stated that the model can interpret sensor data on potential hazards and stop the robot in unsafe conditions.
“The safety system and the functional system need to interact frequently, and in a much broader context. We’ve created this level of operating system and software stack so you can run these two things together,” Goel noted.
Another level of control involves the robot relying not only on its own cameras and sensors but also on the surrounding infrastructure.
Philadelphia-based Fort Robotics is developing controllers and software that gather information from multiple sources. CEO Samuel Reeves explained that it goes beyond merely visually detecting people in the machine’s work zone to more complex data: where a person is, their posture, and whether this information can be trusted enough for the robot to make decisions based on it.
The issue of stability loss has become so sensitive that an expert group from the International Organization for Standardization is specifically studying it. The publication of relevant requirements is expected by mid-2028. In the absence of unified rules, manufacturers are developing their own scenarios.
German company Neura Robotics produces a bipedal robot, 4NE1, weighing about 80 kg. Company founder David Reger stated that the design minimizes risk: if the robot detects a problem, such as a knee joint failure, it will attempt to regain balance, and if unsuccessful, it will fall “like a collapsing building,” folding downward.
Some developers have decided to eliminate the source of risk altogether. Dexmate is creating robots on wheeled platforms with long arms that can reach items on warehouse shelves. Co-founder Yujie Qin explained that the battery and electronics are housed in the platform, giving the machine a low center of gravity and preventing it from falling.
Cobot founder and CEO Brad Porter suggested assessing the threat without exaggeration. His company makes wheeled robots with manipulators that push carts in hospitals or sort parts in factories. They move at walking speed and do not have an overly strong grip.
“We don’t need to put a lot of energy into the necessary actions. We’re not trying to crush watermelons or anything like that,” Porter noted.
In July, Humanoid introduced an approach to training humanoid robots through trial and error on real production tasks.
