Beyond a Single Human Demonstration:
Learning from Scored Trajectory Alternatives for End-to-End Driving

DriveSTA at a glance: supervision paradigms, benchmark comparisons, HUGSIM driving sequences, and a NAVSIM interaction.
Human driving logs record one executed trajectory per scene, while an end-to-end planner must generate and compare multiple feasible choices. DriveSTA turns a single demonstration into scored trajectory alternatives: we construct candidate paths, evaluate their execution, and retain diverse, high-quality targets to supervise planning. A shared Joint-Embedding Predictive Architecture (JEPA) encoder provides efficient temporal context from observed video. DriveSTA reaches 95.1 PDMS on NAVTEST with NAVSIM-only training; DriveSTA-Scale, trained with real and synthetic data, achieves 95.6 PDMS, along with 56.1 route completion and 46.7 HD-Score on HUGSIM.
DriveSTA demosNAVSIM ยท HUGSIM

Driving logs record one executed trajectory, while a planner must generate and compare many feasible alternatives. DriveSTA turns these records into scored trajectory supervision: constrained path construction and screening yield 2.61M high-quality targets from 103.3k real scenes. Dense matching teaches distinct proposals, and replay of successful and failed paths trains their scores. A second stage learns local residual corrections, guided by a pretrained Joint-Embedding Predictive Architecture (JEPA) video encoder adapted to driving observations. DriveSTA-Scale reaches 95.55 NAVTEST PDMS; its Proposal planner scores 91.32 NAVTEST EPDMS and 56.13 NAVHARD EPDMS. On 345 HUGSIM scenarios, Refinement achieves 56.09 route completion and 46.69 HD-Score.

Method

DriveSTA: scored trajectory construction, dense generation supervision, residual refinement, and JEPA-conditioned scoring.

HUGSIM

High-scoring closed-loop episodes ยท Four difficulty levels

Episode RC โ†‘
Episode HD-Score โ†‘

Selected high-score episodes. Aggregate results below cover all 345 scenarios.

Results

Complete evaluations ยท Higher is better
ModelStageNAVTEST
PDMS
NAVTEST
EPDMS
NAVHARD
EPDMS
HUGSIM
RC
HUGSIM
HD-Score
DriveSTA194.7691.1550.8051.8836.88
DriveSTA295.14โ€”โ€”55.2244.05
DriveSTA-Scale195.0991.3256.1354.0543.97
DriveSTA-Scale295.55โ€”โ€”56.0946.69

EPDMS results use Proposal planners with benchmark-specific selection. PDMS: Predictive Driver Model Score; EPDMS: Extended PDMS; RC: route completion. Base: 103.3k real training scenes. Scale: 232.7k real and synthetic scenes.

DriveSTA-Scale Refinement ยท All HUGSIM scenarios by difficulty
DifficultyScenariosRC โ†‘HD-Score โ†‘NC โ†‘DAC โ†‘TTC โ†‘Comfort โ†‘
Easy4984.4775.2293.4098.8791.0792.73
Medium12667.9061.5193.8799.1191.2797.22
Hard8637.8027.6559.3497.3553.1795.23
Extreme8440.5627.3151.3496.6149.3892.69
All34556.0946.6974.8498.0371.5594.98

All scores on a 0โ€“100 scale. NC: no at-fault collision; DAC: drivable-area compliance; TTC: time-to-collision compliance.