Source-linked AI summary
Designing Versatile Samples for Learned Trajectory Scoring
Yaguang Li, Jiaru Zhang, Chuheng Wei, Can Cui, Ziran Wang
TL;DR
The paper addresses limited scorer supervision caused by planners’ concentration around safe trajectories and proposes designed perturbations of logged human trajectories. It trains a transformer scorer over frozen generative planners, achieving 90.1 EPDMS on DiffusionDrive and 90.4 EPDMS on MeanFuser with gains from the designed dataset.
Problem
Strong planners provide few boundary or failure cases in their proposal pools, limiting supervision for learning trajectory-selection decisions.
Method
The method adds lateral and longitudinal perturbations of logged human trajectories and trains a transformer-based scorer over frozen generative planners.
Results
90.1 EPDMS on DiffusionDrive and 90.4 EPDMS on MeanFuser are achieved, with designed-data gains of 0.4 and 0.3 EPDMS respectively.
Takeaways & Limitations
The scorer-training pool is a design variable, and redesigned samples improve PDMS and EPDMS with gains concentrated in product safety terms.
Takeaways & Limitations
The scorer is not uniformly better: drivable-area compliance rises while ego progress and lane keeping fall slightly on both planners.
Abstract
from arXiv · showhide
Many current end-to-end driving policies emit a pool of candidate trajectories and select one, which makes selection a separable component: a scorer can be retrained while the planner, its backbone, and its trajectory generator all stay frozen. However, many strong planners concentrate their proposals around safe mode, providing limited supervision near decision boundaries. In this work, we design a training dataset that provides more informative supervision for the scorer. In particular, we construct two generators that perturb the logged human trajectory along the two axes a vehicle can be displaced: laterally toward the drivable boundary and longitudinally toward a leading vehicle. The designed dataset produces more informative positive and negative samples than the base planner's proposal pool. We attach a transformer-based scorer to two frozen generative planners, DiffusionDrive and MeanFuser, and train it on the NAVSIM navtrain dataset. The results of the experiments show that we achieve 90.1 EPDMS on DiffusionDrive and 90.4 EPDMS on MeanFuser when using ResNet-34, with 0.4 and 0.3 EPDMS respectively, from the designed training dataset.
1. Introduction
Selection-based driving policies separate planning from scoring, but planners’ narrow, mostly safe proposal pools provide limited supervision near decision boundaries. This work augments training with lateral and longitudinal trajectory perturbations and trains a scorer over frozen generative planners.
- Selection-based policies emit candidate trajectories and use a separate scorer to rank them and produce the final output.
- A substantial gap can exist between the selected trajectory and the best trajectory already proposed by the same policy.
- Boundary candidates are rare because competent planners concentrate proposals in a narrow region of trajectory space.
- Binary, cascaded driving-performance components make centimeters of trajectory separation capable of producing opposite targets.
- The method trains a transformer-based scorer on frozen generative planners and adds lateral boundary-directed and longitudinal lead-vehicle-directed samples.
2. Related Work
End-to-end driving has shifted from single trajectories toward generate-select pipelines, while prior work addresses candidate diversity, scoring, perturbation, or simulator-based supervision in different ways. This paper instead trains a standalone scorer over a frozen planner and redesigned trajectory samples.
- End-to-end policies moved from single trajectory regression toward candidate sets because driving is multimodal and alternatives enable assessment.
- Generate-select methods pair a scene-conditioned candidate generator with a scorer that assigns values and selects one trajectory.
- Prior approaches populate decision boundaries through clustered vocabularies, dense candidate sets, multi-head distillation, staged rescoring, or simulator returns.
- This work predicts nine EPDMS components with a standalone scorer while keeping the planner and its perception backbone frozen.
- Trajectory perturbation methods synthesize collisions, off-road excursions, corrective states, or counterfactual transitions to broaden training experience.
3. Preliminaries
NAVSIM evaluates candidate trajectories with EPDMS, whose safety terms multiply and can zero the score after a single violation. The architecture freezes the planner, independently predicts metric components, and composes them for candidate selection.
- NAVSIM candidates contain 8 waypoints of (x, y, θ) over a 4 s horizon sampled at 2 Hz.
- EPDMS combines four rule terms multiplicatively with a weighted average of five quality terms.
- A single violation in the multiplicative block zeros the score regardless of comfort or progress.
- A frozen planner emits candidates, while a scorer ranks them without modifying the planner, backbone, or candidate generator.
- The scorer uses nine independent transformer decoders to predict EPDMS components from scene context and each candidate trajectory.
- Eight binary components use focal BCE, ego progress uses squared error, and undefined labels are masked from the loss.
- At inference, component predictions are composed into EPDMS and the highest-scoring candidate is deployed.
4. Designing Versatile Trajectory Samples
The method builds graded trajectory rungs around the logged human path, perturbing it laterally toward boundaries or longitudinally toward lead vehicles, then solving displacements to reach target margins through simulation.
- Sample construction: The dataset starts from the logged human trajectory and constructs K perturbed copies whose margins target prescribed distances from rule violations.The human margin is signed, with negative values indicating violation, and displacement δ is measured from the human baseline.
- Sample construction: The two generators walk toward decision boundaries along lateral and longitudinal vehicle-displacement axes.The lateral design moves toward the nearest drivable boundary, while the longitudinal design advances along the ego path toward a lead vehicle.
- Longitudinal design: Longitudinal perturbations advance the ego station along its own path toward the lead vehicle while leaving the path geometry unchanged.While the perturbed trajectory remains on the logged path, DAC, LK, and DDC are unchanged by construction; beyond its end, the ego continues along its final heading.
- Lateral design: Lateral perturbations use a smooth arc-length ramp and a scene-level normal, producing a graded offset whose headings are recomputed from the displaced path.The ramp is zero at the ego pose and grows along the path, bending the perturbation into a lane change rather than a uniform sideways shift.
- Target solving: The construction sets target margins, numerically recovers the smallest displacement reaching each target, and emits or masks each rung.The solver evaluates trial displacements, finds target crossings, refines them by bisection, and masks scenes with insufficient planner headroom.
- Target solving: The simulator tracks waypoint references with a batch LQR controller and kinematic bicycle model, so supervision is measured on roll-outs rather than raw waypoints.This introduces a distinction between the designed waypoints and the trajectory actually tracked by NAVSIM.
5. Experiments
On NAVSIM navtest, retraining a scorer with the designed dataset improves frozen planners’ performance, mainly through safety-related components, while some progress and comfort components decline.
- NAVSIM Navtest Result: 90.1 EPDMS for DiffusionDrive and 90.4 EPDMS for MeanFuser show gains from re-ranking frozen planner proposals.The planners, backbones, and candidate generators remain unchanged.
- Ablation Study: 0.4 EPDMS on DiffusionDrive and 0.3 EPDMS on MeanFuser are attributable to designed augmentation beyond the plain scorer.The plain scorer is trained on each planner’s own proposals alone.
- NAVSIM Navtest Result: 3.70% to 3.35% is the reduction in gate-failing deployments across 12,146 navtest scenes.DAC failures decrease from 2.05% to 1.73%, while NC failures decrease from 1.70% to 1.65%.
- Component Score Analysis: Every safety gate improves or holds on both planners, closing 10.0–29.4% of DiffusionDrive headroom and 4.8–26.7% for three MeanFuser gates.Driving-direction compliance is unchanged on MeanFuser.
- Designed Against Random Samples: 89.72 and 90.22 EPDMS for random samples trail the designed set’s 90.05 and 90.41, indicating that sample design matters beyond data quantity.The random and designed sets use matched supervised-row counts.
6. Conclusion
The paper presents a designed training set for learned trajectory scoring, perturbing logged human trajectories laterally and longitudinally and labeling them with scene-specific geometry and benchmark tracking.
- The training set perturbs logged human trajectories laterally and longitudinally in geometry-graded steps.Labels are produced through the benchmark’s own tracker.