This is a general point rather than an argument supporting the GP comment, but it can be shockingly hard to know what the correct response to particular situations is, or predict how a system will respond. For example, one system I know had difficulty differentiating people laying on the road from unmapped speedbumps (particularly at night), but was very good at recognizing them standing up. Someone spent a lot of time testing with dummies until perception improved. Another had no idea how to handle people in fursuits and halloween costumes, because it was incorrectly applying animal behavioral models to pedestrians.