I guess it also depends on the domain and action space. E.g. even in AV I'd expect Gaussians on yaw/vel to be enough, but for juke/jerk we probably want multimodal distributions.

In my (very limited) experience: i) You can get multi-modality on trajectory level even with unimodal Gaussians, so actions being MM isn't as important. ii) When I was using a GMM, nearly all modes collapsed and it was basically a Gaussian + slightly more robust to outliers.