The robotics platform PrismaX points out that a humanoid robot weighing approximately 50 kilograms, powered by lithium batteries, and with a peak output of several horsepower, is essentially heavy machinery; whether or not it incorporates AI does not change its physical risks.
Large models often experience hallucinations or misinterpretations of commands, resulting in erroneous text. Similar problems, when incorporated into embodied AI, can translate into collisions, pinching injuries, falling objects, or even battery accidents. Therefore, embodied AI needs to address two layers of safety simultaneously: the underlying hardware must limit speed, force, and range of motion, and be equipped with independent emergency stops; the upper-level model must be able to recognize uncertainty, refuse dangerous tasks, and promptly hand over control to human intervention.
