To train the GNN with the simulation data on the 3D mesh scene, the simulator data has to be fed into the GNN. Currently, there are some incompatibilities in the way the two deal with data. In particular:
1 - we need to take the parquet file output by the simulator and create a pytorch dataset to be used by the GNN.
2 - to display the tracks predicted by the GNN, we need to take the rollout data from the GNN and create parquet files for input to the simulator trajectory display program.
3 - a couple of places where dimensions are calculated in the GNN assume 2D data.