Why predict the trajectory offset instead of directly predicting the trajectory?

[rogercmq]

Regarding the question mentioned in the title, I found that if we predict the offset given a noisy anchor, the resulting trajectory may appear very zigzag, even for the top-1 trajectory. Direct trajectory prediction, on the other hand, could produce smoother and more coherent paths since it avoids compounding noise from intermediate steps.

[TXYUU]

I also met the problem : pred traj is not smooth, how do you solve it? Or can you share some conclusion ? Thank you very much!

[DhlinV]

I also met the problem : the predicted trajs are all zigzag paths. I guess using plan_anchor or predicting the trajectory directly would address the issue as you mentioned.