About Anchors

[jorgedanielrodrividal]

Dear authors,

Congratulations on your great work.

I have been working on how to generate the plan anchors for my own model and I see two possible approaches:

• Do not process the extracted raw data at all. Take the raw future waypoints from the dataset, recenter them to the ego car, apply your rotation to align with the ego frame and then flatten them into eight-dimensional vectors. KMeans is run directly on these flattened trajectories, and the resulting cluster centers are reshaped into anchors. This approach preserves the raw geometry of the trajectories as they were collected, so the anchors fully reflect the dataset’s inherent info. Because nothing enforces consistency across samples, the spacing between waypoints is irregular, the first waypoint after the origin may not be well aligned, and in some cases the forward component can even move backward. The benefit of this method is that you retain full fidelity to the raw data distribution, so sharp curves or unusual behaviors are represented naturally. The drawback is that the anchors can look messy and unstructured.

• Process the extracted raw data. Instead of clustering the raw waypoints directly, smooth each trajectory using a quadratic or cubic polynomial fit. This fit is explicitly constrained to pass through the origin, and the output waypoints are resampled on a canonical timeline starting at time zero. As a result, every anchor begins at exactly the same origin, and all trajectories share consistent temporal spacing. On top of that, enforce that the forward coordinate never decreases, which guarantees that anchors only move forward in X. Once clustering is done, the anchors are postprocessed to ensure the origin is exact and that forward progression is preserved.

Did you follow any of these two approaches, a combination of both or none of those?

Thank you in advance.