Open-Vocabulary InstructNuPlan

Safety-aware, open-vocabulary driving instruction annotations over nuPlan scenarios, from the WACV 2026 paper iMotion-LLM: Instruction-Conditioned Trajectory Generation.

This dataset contains annotations only — no nuPlan sensor data, maps or logs. Every record is keyed to a nuPlan scenario token. To use it you must obtain nuPlan yourself from https://www.nuscenes.org/nuplan under Motional's own license and terms.

Why this exists

The instruction/reasoning pairs here are LLM-generated output. They were produced by prompting a GPT model once, per scenario, and are therefore not reproducible by re-running the pipeline: generation is non-deterministic and costs API credits. Publishing them is the only way to reproduce the paper's open-vocabulary numbers exactly.

Everything else in the paper's preprocessing chain is reproducible from code and is deliberately not published here — see "What is not here".

Coverage

14 nuPlan scenario types:

accelerating_at_crosswalk, behind_bike, behind_long_vehicle, following_lane_with_lead, following_lane_without_lead, following_lane_with_slow_lead, starting_protected_cross_turn, starting_protected_noncross_turn, starting_unprotected_cross_turn, starting_unprotected_noncross_turn, stopping_with_lead, traversing_crosswalk, traversing_intersection, waiting_for_pedestrian_to_cross

Configs

Config	Rows	What it is
gpt_responses	7,119	The core dataset. Per scenario: the prompt sent to the LLM and its returned instructions, each with reasoning, safe / safe_instruction flags and a category (e.g. "Without Context").
template_instructions	88,133	Template-rendered instruction/reasoning/decision records used to condition the trajectory model (Instruction, Reasoning, Decision, Label, Direction_cls).
meta_prompts	7,131	Structured scenario metadata: scenario_type, situation, what_safe_to_do, what_unsafe_to_do.
positive_prompts / negative_prompts	7,131 each	Raw prompt text for the safe / unsafe instruction arms, kept for provenance.

Every row carries _scene_id (nuPlan token), _source (original relative path) and, where a source file held several records, _line.

Loading

from datasets import load_dataset

ds = load_dataset("KAUST/InstructNuPlan", "gpt_responses", split="train")
print(ds[0]["content"][0]["data"][0]["instructions"][0])

Format note

The original release was ~43,000 small .txt / .json files. They are packed here as gzipped JSONL so the dataset loads in one pass instead of tens of thousands of HTTP requests. _source preserves the original path of every record, so the on-disk layout is fully recoverable.

What is not here (reproducible from code)

Deliberately omitted, because the released code regenerates it from nuPlan:

• npz scenario tensor caches (~912 GiB)
• rendered scenario figures (.png)

See the code repository for the preprocessing and prompt-construction scripts.

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 nuPlan data remains subject to Motional's license, which you must accept separately. Redistribution of nuPlan sensor data, maps or logs is not granted by this repository.