AV safety in edge cases is important.
We finally have a reliable way to benchmark and compare these systems when routine displacement metrics are not enough.
FluidTestNext Challenge Coming at July
AV safety in edge cases needs a reliable benchmark.
FluidTest turns long-tail driving scenes into comparable planner safety evidence, so model teams can see whether their systems introduce additional threats beyond human driving logs.
Long-Tailwith Challenging Scenarios
HumanPairwise Safety Preference
AgenticExplainable and Auditable
FluidTest benchmark
Edge-case safety is finally comparable.
Reliable AV evaluation starts where routine metrics are weakest: rare scenes, planner disagreements, and behavior that can introduce new threats beyond the human driving log.
Planner outputs are judged against human logs.
Labelers compare each planner result against the human driver's log and judge whether it brings additional threats.
Teams can compete on validated safety.
Submissions are validated, labeled, reviewed by a neuro-symbolic agentic system, and ranked on the leaderboard.
How it works
Labelers compare planner behavior against the human driver's log.
Each panel pairs a projected planner path with the threat class a labeler can review before the final neuro-symbolic audit. The red trajectory is the planner's result, and the cyan trajectory is the expert reference trajectory.Right-of-way projection
Vehicle yield conflict
Projected ego motion enters another road user's priority path instead of holding behind the human log.S2 yield conflictPoutineSFT parked-vehicle projection
Stationary vehicle collision
Projected ego occupancy overlaps a parked vehicle while the human log gives it clearance.S3 collision coursePoutineSFT signal approach
Yellow-phase noncompliance
The planner proceeds into a safely stoppable amber phase while the human log remains stopped.S2 signal conflictRoadway position
Off-road driving
Planner trajectory leaves the normal lane and enters a gore, separator, shoulder, or curbside surface.S1 departurePoutineSFT corridor drift
Meaningless lateral drift
The predicted path deviates laterally from the human log inside a narrow corridor without a visible reason.S1 lateral driftJoin as a labeler
Push AV systems forward with careful human review.
Register, learn the review protocol, pass the test, earn a certificate, and contribute labels that shape how autonomous planners are evaluated.
- 01Register
- 02Learn and test
- 03Get certified
- 04Contribute labels
For AV teams
Warm up on the FluidTest leaderboard.
FluidTest leaderboard is currently in warmup. To try the baseline, download Val151, run UniPlan inference, and upload the result to preflight; when the projection looks correct, that baseline pipeline is complete. To beat the baseline, train on WOD-E2E, evaluate ADE, infer on Val151, review the projection, then submit an official result for human and agentic review.
Baseline Instructions
Download Val151, run the UniPlan warmup baseline, and upload the predictions for a projection check.
->Unlimited Preflight Checks
Use preflight as often as needed to validate JSON format, Val151 coverage, and projected trajectories.
->Official Submission Limit
After projection review, each account and team can submit up to 3 official leaderboard results per month.
->Private Until Released
Track previous submissions and choose public or private visibility after final review is ready.
->