nuplan-devkit
nuplan-devkit copied to clipboard
nuPlan paper table comparison
Reposting here as suggested in (https://forum.nuscenes.org/t/nuplan-paper-comparison-table/797):
Hello ! I was reading the nuPlan paper (https://arxiv.org/abs/2106.11810) where the Table 1 (shown below):
Is showing that the Lyft dataset doesn't have Closed-loop evaluation and that it has a "prediction" in the Type column.
As a suggestion, I would review the table because in L5kit there is a closed-loop evaluation with many closed-loop metrics available since 2021 (before the nuPlan paper publication). Not only there is a simulator but also a RL environment where rollouts can be done and use rewards from closed-loop metrics as well. There is also an entire section about Planning where there was even a paper released using it for training a planner in closed-loop. Anyway, the table doesn't seem to reflect the current state of the framework.
Hi @perone,
Sorry for the super late reply. Thank you for pointing this out. We are in the process of reviewing the paper to make it more accurate
No problem, thanks a lot for the reply @patk-motional.
Hi @patk-motional, can we reopen this issue ? It has been more than 1 year that the paper doesn't reflect the correct information about the framework and other datasets, that doesn't seem a fair comparison.
Hi @perone,
Sorry for the delay, does something like this suffice?
Thanks for the quick reply @patk-motional, this sounds fine. I think the type can also be Pred+Plan (also for nuPlan no ?) because it can be used for agent prediction tasks as well, but up to you.
After looking into this we have decided to update the paper to state that your dataset provides a closed-loop planning tutorial, but no official benchmark, which is really the emphasize of nuPlan.