Cost, reliability, and the dexterity problem
Three things generally have to improve for humanoid robots to reach real mass deployment. First, cost has to come down enough that a robot competes with the fully loaded cost of human labor for a given task in a given market — which varies a lot by country and industry, meaning humanoids may make economic sense in some markets and tasks well before others. Second, reliability and uptime have to reach a level where a robot can be trusted to work through a shift without constant human supervision, intervention, or resets, since unreliable robots create more work monitoring and fixing them than they save.
Third, and often described as the hardest unsolved piece, is genuinely general manipulation dexterity — reliably picking up, handling, and manipulating a wide variety of objects with human-like hands, including objects that are soft, oddly shaped, slippery, or that the robot has never encountered before. Locomotion, meaning walking, balancing, and recovering from being pushed, has improved dramatically across the industry in recent years and is reasonably well understood by comparison. Manipulation and dexterity have not caught up nearly as far, and it remains the harder engineering and AI problem standing between today's narrow deployments and truly general-purpose humanoid labor.