Certifying a collaborative humanoid for a working factory floor is not the same problem as certifying a fixed industrial arm. The robot shares space with people, moves through an unstructured environment, and is expected to behave predictably in situations no one fully enumerated at design time. These notes capture what stood out across the engagement.
01The hazard analysis has to assume a person is always nearby
In a caged workcell you can engineer the human out of the hazard zone. On a shared factory floor you cannot. Every hazard analysis assumed continuous human proximity, which moved the burden onto the robot’s own safety functions — force and speed limiting, safe stopping, and reliable detection of people in the path.
02Moves like a person, fails like a machine
A humanoid form invites people to read human intent into the machine’s motion. That is exactly the trap to avoid in a safety argument. The validation had to treat the robot as a machine with quantifiable failure modes, not as a colleague — separating the ergonomics of collaboration from the cold accounting of what happens when a sensor, actuator, or controller fails.
The throughline is the one we return to on every program: independence and evidence beat intuition. A humanoid that looks safe is not the same as one whose safety you can prove.