Heterogeneous model setups support

Are there any plans to support heterogeneous scenes, where you can swap grippers/objects across envs without having to recompile the model? It seems to be a big bottleneck in playground manipulation vision-based environments that heavily relies on augmentation for sim2real transfer. disabling constrains of unused objects through eq_active doesn't seem to be effective at scale (trying to augment 1,000+ objects).