Input data != target data for LSTM

Hello there!
I have some doubts about LSTMs and how to preserve their temporality. Let me explain a little bit:

I have as input a trajectory of 12 positions with different information: my input data is (batch_size, 12, 110) where 110 are 110 columns with different information about that position.
As output/target I have the lat/lon coordinate: target/output data is (batch_size, 12, 2) where 2 represents the latitude and longitude of the position.

Once trained the network I represent the coordenates into the map and what I see is that the sequential benefit of LSTM is not being taken advantage of.
I’ve thought about teacher forcing but before that I wanted to know if there’s a way of exploiting sequential information this way in LSTM architecture… Any idea?

Thanks!!