The major benefit is that human behavior capture trajectories can be immediately mapped to robotic joint control points.

The goal is to test what specifically unique to bipedal humanoid action space that we can’t with something more stable.

My personal preference for robot form factor is a centaur, but nobody makes those at scale unfortunately

Seems to me, any case where a robot needs to navigate an environment built for able bodied humans. Quadrupeds get close, but even they will encounter spaces we have organized for ourselves that they can't navigate in. An immediately useful and saleable use case would be a two legged robotic exoskeleton to replace wheelchairs or mobility scooters. More than replace, being able to perform actions like kneeling down to do gardening or play with kids. Or why not go superhuman and have extendible legs letting you reach the top shelf at the supermarket. Give the robot an opaque skin with open able flaps and the operator could use a regular lavatory for a bit more human dignity.