NVIDIA DGX Spark
Awarded to the Track 1 winner. Personal AI supercomputer for desktop development and inference.
Autonomous driving stacks behave well on the head of the distribution, but the long tail — rare, ambiguous, or interactive scenarios — is where most safety-critical failures occur. This challenge invites the research community to build models that can reason about these long-tail scenarios in natural language, building on NVIDIA's public PAI-AV Dataset.
Models will be evaluated on a curated out-of-distribution test set mined from a large physical-AI autonomous-driving corpus. Each scenario is anchored at a precise keyframe and annotated with a chain-of-causation describing the relevant agents, interactions, and the appropriate ego behavior.
The 2026 edition has two tracks: a chain-of-causation reasoning-generation track, and an open auto-labeling leaderboard for the research community.
Input: a multi-camera driving clip and an event window.
Output: a free-form natural-language explanation that identifies
the relevant agents, the interactions that make the scenario challenging,
and the recommended ego behavior at the keyframe.
Input: the validation clips of the out-of-distribution reasoning set.
Output: automatically generated chain-of-causation reasoning labels
for each clip.
Submit a single submission.json in the Submit tab of the
leaderboard below. Evaluation runs automatically on GPU and your scores return
in minutes. Untick "Publish to public leaderboard" to receive your scores
privately without adding a row to the public board. Every submission maps each
test sample key "<clip_id>_<event_idx>" (event index
within a clip, starting at 0) to a list of reasoning rollouts; missing samples
are scored as 0, so omitting hard cases cannot help.
Format: 6 rollouts per sample.
{
"<clip_id>_<event_idx>": [
"rollout 1 reasoning …",
"rollout 2 reasoning …",
"… up to 6 rollouts"
]
}
Scoring: each rollout is scored against the hidden ground-truth reasoning by the AlpaJudge alignment model (probability in [0, 1]). Reported per sample, then meaned over the test set:
Getting the test data: download the Track 1 test parquet —
ood_reasoning_test.parquet
(214 clips · 284 events). It is indexed by clip_id; the
events column is a JSON list of
{"event_start_frame", "event_start_timestamp"} — the
ground-truth chain-of-causation labels are withheld. Build one submission
key per event, where event_idx is the event's 0-based
position in the list:
import pandas as pd, json
df = pd.read_parquet("ood_reasoning_test.parquet")
keys = [f"{clip_id}_{i}"
for clip_id, row in df.iterrows()
for i, _ in enumerate(json.loads(row["events"]))]
Format: 1 rollout per sample.
{
"<clip_id>_<event_idx>": ["your single auto-generated reasoning label"]
}
Scoring: the top1 AlpaJudge score only (ranking metric), meaned over the test set.
Getting the test data: the Track 2 test set is the
validation split (split == "val", 290 clips) of
reasoning/ood_reasoning.parquet
in the PAI-AV dataset (gated — accept the dataset license and log in with
your HF token first):
from huggingface_hub import hf_hub_download
import pandas as pd
path = hf_hub_download("nvidia/PhysicalAI-Autonomous-Vehicles",
"reasoning/ood_reasoning.parquet", repo_type="dataset")
val = pd.read_parquet(path).query('split == "val"')
As in Track 1, submission keys are
"<clip_id>_<event_idx>" with
event_idx the event's 0-based position in the clip's
events list.
Trouble loading? Open the leaderboard directly: huggingface.co/spaces/nvidia/PhysicalAI-OOD-Leaderboard ↗
Dates are tentative and subject to update.
The winner of Track 1 will be awarded an NVIDIA DGX Spark. Track 2 runs as a leaderboard-only benchmark for the community.
Awarded to the Track 1 winner. Personal AI supercomputer for desktop development and inference.
This competition is hosted by NVIDIA's Autonomous Vehicle Research Group.
HostInterdisciplinary NVIDIA Research team advancing vehicle autonomy across perception, prediction, planning, control, simulation, foundation models, and AI safety.
Bookmark this page — full challenge details, evaluation methodology, and submission instructions will be published when the evaluation server opens.