states.learned

DETR-based trackers (MOTR, TrackFormer, MeMOTR) do not filter the track embedding with a hand-written motion model; they let a *learned* module propagate the track query from frame to frame and fuse the new detection. That is a learned recurrent filter, not a closed-form one, so unitrack supports it as a pair of hooks that wrap any callable / ``torch.nn.Module``: - :class:`LearnedProcess` is the predict step — ``query <- module(query, dt)``. - :class:`LearnedObservation` is the update step — ``query <- module(query, measurement)`` for matched pairs. Because the wrapped module is autograd-native, these are the differentiable member of the embedding-filter family: a tracking loss can backpropagate into the propagation/update module (and through it, the backbone).