Nice, thanks paroj, I removed that logic from the custom importer and it behaves the same, so at least, the code is much simpler now...
I've been able to replicate the same result as the FBX importer from Blender, so IDK if it counts as a win... haha
Jokes apart, on the custom importer, it looks good now, but the T-pose stands on the Z axis while the animation happens standing on the y axis...
I thought it could be because I'm not applying the root transformation when processing the mesh, when I did it, it looked exactly as it does in Blender.
I also tried to load it in the Autodesk FBX Review, there it is in the y coordinate as well, but even the T pose is in the same axis.
When loading using the Assimp plugin, it gets that weird animation, but at least the T-pose and the animation are oriented on the same axis.
edit: It might be related to this as well: in the custom importer, I'm building up the skeleton, calculating the bones' orientation based on the parent bone, when I do this, I get the global transformation by getting the inverse of the matrix, and then the local considering the parent bone global transform.
But in the plugin it calculated all the transformations using only the Assimp mTransformation, and then, in the end, it got the inverse();
Anyways, I need to investigate it further, I'll wrap up the custom importer, it got better now, and then look deep into this and the plugin to try to fix and get the same behavior on both importers... at least this robot is a good challenge