Training-free · Test-time optimization

DriveTTO:Understanding Search and Selection
in Test-Time Trajectory Optimization

More from a frozen driving planner.

Search around an existing trajectory. Compare complete candidates.
Improve the plan at inference time, with the planner held fixed.

See DriveTTO in action6 min 36 s · Endpoint search · Refinement · Driving
Download introduction
NAVTEST

95.5PDMS ↑

NAVSIM-v1
NAVTEST

92.3EPDMS ↑

NAVSIM-v2
NAVHARD

57.1EPDMS ↑

Two-stage evaluation
HUGSIM

54.5RC ↑

345 scenarios
HUGSIM

40.1HD-Score ↑

345 scenarios

Paper benchmark results. Each benchmark uses its reported configuration; protocol details accompany the tables.

01 / OVERVIEW

A better plan.
The same planner.

A frozen driving planner still leaves room to improve its trajectory at inference time.

DriveTTO searches around the planner’s output, proposing complete trajectories through endpoint search and temporal fusion. It scores the candidates alongside a preserved reference and returns the selected trajectory, without updating the planner’s parameters.

We examine this approach in NAVSIM evaluation and HUGSIM closed-loop driving. The results distinguish the full strategy, search variants, and benchmark-specific configurations, together with their runtime and transfer behavior.

Frozen plannerTrajectory searchSelected plan

02 / METHOD

Search the trajectory.
Keep the model fixed.

Construct complete alternatives around a retained plan, then compare candidates and the reference using frozen scorers.

ANIMATED EXPLAINER

From endpoint search to precise motion.

A continuous 2-minute 48-second illustration of endpoint feedback and two-planner refinement.

Watch complete-path evaluations reveal a promising endpoint region, then follow the two planners through local refinement and matched-time fusion. Close-ups show how small position changes affect vehicle clearance. Geometry and utility are schematic.

Download animation ↗ Explore interactive animation ↗
DriveTTO / Method overviewClick to inspect the complete figure
  1. 01

    Preserve the reference

    Start from the frozen planner’s output and retain the reference trajectory in the candidate set.

  2. 02

    Propose complete trajectories

    Explore endpoint variations and temporal fusion to construct candidate trajectories around the existing plan.

  3. 03

    Compare and return

    Evaluate complete candidates with the configured scoring rule and return the selected trajectory for execution.

03 / CLOSED-LOOP DRIVING

On the road,
scene by scene.

Inspect selected HUGSIM rollouts with their scene conditions and evaluation metrics.

Select a rollout below.

Selected examples illustrate individual rollouts. Aggregate performance is reported in the benchmark tables below.

04 / PLANNING VISUALIZATIONS

The plan, in context.

Illustrative NAVSIM examples, selected at 100 PDMS. Switch between the front and bird’s-eye views, or open a figure to inspect the full image.

Loading planning visualizations…

Plans in motion

Recorded front-camera frames and calibrated plan projections, paired with a bird’s-eye view. These open-loop animations follow the saved Full / seed 0 plan; camera motion and other agents come from the recorded scene.

05 / EXPERIMENTS

Results, with context.

Compare the full strategy, search variants, and planner transfer under their reported protocols.

NAVSIM benchmark comparisons

Paper-reported scores across three evaluation settings. Higher is better.

Manuscript comparison rows

Full component breakdown Scroll horizontally on smaller screens.

Comparison rows are transcribed from the manuscript’s benchmark tables and retain their reported configurations. They are distinct from the controlled, same-resource experiment below.

Controlled comparison · three seeds, fixed model resources

12,146 observed NAVTEST scenes, 44 parent logs, three search seeds. Full reaches 95.5116 PDMS, improving over CEM8 by 0.4075 (95% CI [0.1149, 0.7274]; Holm-adjusted p = .036).

Isolated p50 latency is 1.3093 s for Full and 1.1998 s for CEM8 (+9.1%). All timings include the common dual-model encoding; the Native row is not a standalone single-model deployment latency.

After averaging seeds within scenes, drops exceeding 20 points from Native occur in 118/12,146 Full scenes and 11/12,146 CEM8 scenes. Minimum, mean and DrivoR terminal ranking show no resolved mean-score difference; this comparison does not isolate temporal fusion.

Inspect the numbers behind the page.

Download results CSV

06 / RESOURCES

Explore DriveTTO.

Code, reported results, and selected visual evidence in one place.

Paper coming soon. A paper link and citation will be added when available.

Planning visualization