Welcome to the nuReasoning Challenge 2026. The challenge asks two things on every held-out test clip: plan a 5-second ego trajectory from the key frame, and reason about the scene by answering multiple-choice questions grounded in the same moment.
nuReasoning is a reasoning-centric multimodal benchmark for end-to-end autonomous driving in long-tail scenarios: 20K curated ~20-second clips at 10 Hz with synchronized 8-camera images, LiDAR, HD maps, ego states, object annotations, and frame-level Spatial, Driving, and Counterfactual reasoning annotations. The private test split holds 1,000 clips and 3,000 questions; participants receive cameras, ego history through the key frame, and the questions without answers or future states.
You run your model locally, generate one challenge_submission.json with the devkit, and upload it here. Trajectories are scored with the nuReasoning Planning Score (NPS); answers are scored by exact multiple-choice accuracy.
| Metric | Weight | Definition |
|---|---|---|
planning_score |
0.75 | Mean nuReasoning Planning Score (NPS) over 1,000 clips. NPS gates on no-at-fault collision and drivable-area compliance, then combines progress (0.3), comfort (0.2), and human-likeness (0.5). Missing or invalid trajectories score 0. |
reasoning_score |
0.25 | Exact multiple-choice accuracy over all 3,000 expected questions. Missing, duplicate, or invalid answers score 0. |
final_score |
1.00 | 0.75 × planning_score + 0.25 × reasoning_score. Ranks the
leaderboard; higher is better. |
The submitted target_frame_index must match the clip key frame (100 on this test
set) or that clip's planning score is zeroed. Planning, reasoning, and final scores are all
shown on the leaderboard.
51 waypoints [dx, dy, dtheta] at 0.1 s over a 5 s horizon, in the key-frame ego frame. Scored with NPS against held-out futures and maps.
Spatial, Driving, and Counterfactual multiple-choice questions at the key frame. One answer A–D per question.
Both tasks are submitted together in a single file and ranked by a single final_score. There is one leaderboard.
The current phase of the challenge and the evaluation configuration in effect. Material changes to the metric, data, or rules will be recorded here.
No known issues in the current evaluator. If a submission that validates locally is rejected or scored unexpectedly, open an issue on the devkit GitHub with your submission ID.
Every team must register before its results count. Registration is free and takes a few minutes; it lets the organizers match leaderboard entries to real teams, enforce the one-team rule, and contact prize winners.
Provide your team name, affiliation, team members, a contact email, and the Hugging Face username(s) that will upload submissions.
Use the same Hugging Face account you listed in the form. A team entry is created on your first visit to the Submit tab.
On the Submit tab, set your team name to exactly the team name from the form. This is the name shown on the leaderboard.
Consistency requirement. Your challenge entry must be consistent with your registration: the leaderboard team name, the Hugging Face account(s) that submit, and the affiliation must match what you registered. Entries that cannot be matched to a registration are not eligible for prizes.
The challenge compares independently developed planning-and-reasoning systems under a shared evaluation protocol. These rules protect that comparison.
Each participant may belong to only one team, and each team uses one Hugging Face account (or the accounts listed at registration) to submit. Do not create multiple accounts to exceed the daily limit. Team mergers require organizer approval and an updated registration.
Leaderboard entries must match the registration form: team name, submitting account(s), and affiliation. Unregistered or mismatched entries may be removed and are ineligible for prizes.
You may use the released nuReasoning train and validation splits. Pretraining on other public datasets and using public foundation models is permitted. Do not train, tune, or otherwise use the private test labels, future trajectories, or maps.
Do not attempt to recover, scrape, or reverse-engineer private labels, future trajectories, or maps. Manual labeling of test clips is prohibited. Submissions that show evidence of label leakage will be removed.
Submit the official per-clip JSON described on the Submit tab, generated and validated with the devkit. Up to 3 submissions per team per day.
Enter a method name (optionally hyperlinked to a paper or repository) with each submission. Prize candidates may be asked for a short technical description and reproducibility evidence; non-reproducible results can be invalidated.
Follow the Hugging Face Hub terms of service and the license on the nuReasoning dataset card. Organizers and individuals with access to the private labels may submit for reference but are not eligible for prizes.
Decisions on scoring bugs, disqualifications, ties, and late submissions are final. Clarifications are posted in the Competition Status section above.
All times are UTC.
A total prize pool of USD $10,000 is awarded across three prizes. Prize eligibility requires a valid registration consistent with the leaderboard entry and compliance with the rules.
The team ranked first on the final leaderboard by final_score.
The team ranked second on the final leaderboard by final_score.
Awarded to the team whose approach the committee judges most novel or insightful for planning and reasoning. May overlap with a placement prize.
Everything you need to download the data, train a baseline, and generate a valid submission lives in the open-source devkit.
Dataset download and visualization, nuVLA training, planning benchmark, reasoning fine-tuning, and the challenge submission tool.
Gated Hugging Face dataset with the train and validation splits. Request access, then download with the devkit.
Environment setup, dependencies, and dataset download instructions.
How to run nureasoning.submission.challenge, the JSON schema, and local
validation before upload.
Reference planners and the nuVLA training recipe to get a first score on the board.
VQA generation, LoRA fine-tuning, serving, and evaluation for the multiple-choice questions.
Benchmark overview, annotation schema, and the nuReasoning paper.
Use GitHub issues for bug reports, scoring questions, and rule clarifications.
The nuReasoning Challenge is jointly developed by Motional and the UCLA Mobility Lab.
Driverless technology company formed by Hyundai Motor Group and Aptiv, with roots in nuTonomy and the creators of the nuScenes and nuPlan benchmarks.
UCLA Mobility LabResearch lab at the University of California, Los Angeles working on cooperative driving automation, perception, planning, and open-source autonomy tooling.