@bobthebuilder: I have posted an update on this problem here : [Important] Dataset + Problem Update
Thanks for all your inputs.
We also took the time to simplify the problem a bit more.
We could have temporarily solved the problem by ensuring that Gimbal Locks are not represented in the training and test sets. But on some initial exploration, we realised that most models would still struggle to learn properly in that loss landscape. Our whole goal to use Euler Angles was to make the notion of orientation intuitive for the participants. And trying to bring in Quaternions now would solve the problems here, but also increase the complexity of the task from a ML pov. So we would leave that for a future version of AIcrowd Blitz. In the meantime, we come up with a simplified problem statement, where the goal is to predict the orientation around an arbitrary axis that we used to generate the training and test set.
Cheers,
Mohanty