Why do we need to process Waymo files?

GPUDrive works with a subset of information given in a Waymo scene, in json format.
Processing script deserializes Waymo tfrecords according to their protobuf format and generates json files compatible with GPUDrive.

What exactly happens when processing a Waymo tfrecord?

Each Waymo tfrecord contains about 500 scenes, with all assiciated information regarding road objects, vehicles, traffic lights, etc.
process_waymo_files.py generates a json file for each scene in the tfrecord in parallel.
During processing, we mark certain vehicles as "expert" meaning they cannot be controlled in the sim.

Why do we mark some vehicles as experts?

In the ground truth Waymo vehicle trajectories, a small minority of them involve crossing a road edge entity.
GPUDrive interprets this as a collision and subsequently going offroad. Hence these vehicles fail to make it to their goals.
This leads to inaccuracies in evaluating policies with respect to rate of reaching goals.
Marking them as "expert" makes them uncontrolled in the sim, and hence are not considered in evaluation.

How do we check if a vehicle should be marked expert?

For each Waymo scene we process, we construct all road edges, and all vehicle trajectories.
For each vehicle trajectory, if it intersects with a road edge, we set an expert flag in the json.
Then when a scene is loaded into the sim, it checks for this flag and behaves accordingly.

How many of these experts even exist in the first place?

In the entire training set (>100k scenes), we found 31837 vehicles (roughly 0.35% of all vehicles) marked expert.

NuScenes adapter

GPUDrive still consumes the same JSON schema as the Waymo converter. For NuScenes, use the adapter to convert keyframe annotations and map layers into that schema:

python data_utils/process_nuscenes_files.py \
  --dataroot /path/to/nuscenes \
  --version v1.0-mini \
  --split mini_train \
  --output-dir data/processed/nuscenes

The converter synthesizes an ego vehicle from NuScenes ego poses, maps supported dynamic categories to vehicle, pedestrian, and cyclist, and interpolates 2 Hz keyframe tracks to GPUDrive's default 10 Hz, 91-step trajectory format. The default output prefix is tfrecord-nuscenes, so SceneDataLoader can discover the files with its default file_prefix="tfrecord".