What still takes the most engineering time when adding a new Isaac Lab task?

[Jaiparmar940]

For people using Isaac Lab on real robot-learning projects: once the scene, assets, and manager-based or direct task scaffolding are working, what still takes the most engineering time when you add a new training or evaluation task?

Is it validating USD assets and contact dynamics, implementing observations/actions/rewards/terminations, building scenario coverage and domain randomization, checking reward and success-condition correctness, sim-to-real validation, or something else?

I recently built a small open-source vehicle-fault decision environment (https://github.com/Jaiparmar940/rlenv) and am trying to understand where current tooling actually stops. Concrete examples from the last task you completed would be especially helpful: what Isaac Lab handled, and what your team still had to implement or verify.