Losses and norms for a single batch of 50. Looks like it is fairly stable and the learning rate is actually too restrictive. The predicted trajectory were a bit off.