For forty years, industrial robot safety rested on a simple idea: keep people and robots apart. Fences, light curtains and interlocked doors made the cell a place where the robot could move fast because no one was inside. Collaborative robots relaxed that rule by limiting speed and force. Humanoids break it altogether. Their whole value is that they work where people work — aisles, stairs, workstations designed for the human body — so safety has to move with them.
The first difference is physical. A conventional robot arm that loses power simply stops. A humanoid is dynamically stable: it stays upright only because its controllers keep balancing it, so a power loss, a slip or a software fault can become a fall. A 60–70 kg machine falling near a worker is a hazard no traditional standard was written for. Safe behaviour therefore includes controlled fall strategies, a defined fall zone around the robot, and safe states such as crouching or sitting down that the robot can reach before cutting power. Even routine maintenance changes: you cannot lock out a humanoid the way you lock out a press without first deciding where and how it comes to rest.
The second difference is the brain. Many humanoids are driven by AI models — vision-language-action policies trained on demonstrations — whose behaviour is learned rather than programmed line by line. That is what makes them adaptable, and it is also why the AI cannot be the safety system. The robust pattern is layered: the AI proposes motions, while an independent, certified safety controller enforces speed limits, force limits, protective stops and separation distances whatever the model decides. Platforms like NVIDIA Halos push this separation down to the silicon, and add outside-in safety: fixed cameras that watch the area and slow or stop the robot when a person enters its zone.
The standards are catching up, but they are not finished. Today a humanoid deployment is assessed against a stack of existing documents: ISO 12100 for risk assessment, ISO 10218-1 and -2 (revised in 2025) for industrial robots, ISO/TS 15066 for collaborative operation, mobile-robot standards such as ISO 3691-4 and ANSI/A3 R15.08, and ISO 13849-1 or IEC 62061 for the safety-related control system. The first standard aimed squarely at this class of machine, ISO 25785-1 on industrial mobile robots with actively controlled stability, is still in development. In Europe the regulatory frame is also moving: the new Machinery Regulation (EU) 2023/1230 applies from 20 January 2027 and explicitly covers machinery with self-evolving, machine-learning behaviour in safety functions, which requires third-party conformity assessment, as well as protection against cyber-corruption. Under the AI Act, AI used as a safety component of such products is high-risk, with obligations now due from 2 August 2028.
Two risks are easy to underestimate. The first is cybersecurity: a compromised humanoid is not a data breach, it is a moving machine under someone else's control. Network segmentation, signed software updates, secure remote access and an IEC 62443 approach to the industrial network are safety measures, not IT extras. The second is privacy. A humanoid is a mobile platform full of cameras and microphones that records the people it works beside. That is personal data under the GDPR, and in Italy the use of devices that can monitor workers also falls under Article 4 of the Workers' Statute, which requires a union agreement or labour inspectorate authorisation. Process on the edge, keep only what is needed, and involve the DPO and the workers' representatives early.
In practice, a safe deployment starts small and specific: one application, a documented risk assessment, clear zones where humans and robots mix or do not, emergency stops reachable by every operator — including wireless ones — safe charging areas for high-capacity batteries, training for the people who will work alongside the robot, and a monitoring plan that tracks near misses, not just accidents. The employer's risk assessment document must be updated before the robot starts, not after. At Svapna we work on the parts of this picture that are software: computer vision and robotics, real-time video analytics for zoning and alerting, edge inference and secure integration with the systems that already run your plant. Humanoids are coming to the shop floor; the companies that benefit first will be the ones that engineered safety before the robot arrived.