InstructWaymo

Instruction-conditioned trajectory annotations over the Waymo Open Motion Dataset, from the WACV 2026 paper iMotion-LLM: Instruction-Conditioned Trajectory Generation.

This dataset contains annotations only — no Waymo scenario or sensor data. Every record is keyed to a Waymo scenario id. To use it you must obtain the Waymo Open Motion Dataset yourself from https://waymo.com/open/download under Waymo's own license and terms.

Contents

Config	Rows	Source files	What it is
train_instructions	45,011,552	3,616,562	Training instructions: Instruction, Reasoning, Decision (<Accepted> / rejected), Label (gt / contrastive) and Direction_cls (0–7)
val_instructions	359,308	28,932	The same, for the validation split
train_synth_instructions	910,782	303,595	Synthetic/augmented instruction variants
gf_mtr_mapping	5,000	5,000	GameFormer ↔ MTR scenario index mappings

Every row carries _scene_id (Waymo scenario id), _source (original relative path) and _line (position within its source file).

Evaluation manifests

eval_meta/ holds the exact manifests used for the paper's evaluation, as plain JSON:

Path	Contents
eval_meta/validation_30sep/{meta_gt1,meta_pos1,meta_neg1}.json	the ground-truth, positive and negative instruction evaluation lists
eval_meta/validation_30sep/eval_meta.json	full evaluation metadata
eval_meta/feb16_2025_validation/…	the same for the Feb-2025 validation split
eval_meta/agent_lists/actual_agent{0,1}.json	per-agent evaluation lists

These define the splits the reported numbers were computed on and are not recoverable by re-running preprocessing, so they are published verbatim.

Direction classes

Direction_cls indexes:

0 stationary · 1 move straight · 2 move straight veering right · 3 move straight veering left · 4 turn right · 5 turn left · 6 take right U-turn · 7 take left U-turn

Loading

from datasets import load_dataset

ds = load_dataset("KAUST/InstructWaymo", "val_instructions", split="train")
print(ds[0]["Instruction"], "|", ds[0]["Direction_cls"])

To restore the original per-scene file tree the release configs read:

pip install huggingface_hub
python tools/instruct_data/download_instruct_data.py --dest data/processed --expand

(from the code repository)

Format note

The research pipeline produced 3.65 million small .txt files. Packed here as 26 gzipped JSONL shards (0.61 GiB total, from 26.6 GiB of raw text), so the dataset loads in one pass instead of millions of HTTP requests. _source preserves every original path, and --expand rebuilds the tree.

What is not here (reproducible from code)

Deliberately omitted — the released code regenerates these from the raw Waymo data:

• npz scenario tensor caches (~432 GiB)
• per-scene agent/map JSON and acts pickles
• rendered scenario figures (~22 GiB)

Citation

@inproceedings{Felemban_2026_WACV,
  author    = {Felemban, Abdulwahab and Hroub, Nussair and Ding, Jian and Abdelrahman, Eslam and Shen, Xiaoqian and Mohamed, Abduallah and Elhoseiny, Mohamed},
  title     = {iMotion-LLM: Instruction-Conditioned Trajectory Generation},
  booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
  month     = {March},
  year      = {2026},
  pages     = {2710-2720}
}

Licensing

These annotations are released for research use as a derived work. The underlying Waymo Open Motion Dataset remains subject to Waymo's license, which you must accept separately. Redistribution of Waymo scenario or sensor data is not granted by this repository.