World models support autonomous driving by evaluating candidate trajectories through predicted future scene evolution. However, accurate future prediction does not ensure road-compliant execution; when all candidates violate road boundaries, re-ranking cannot select a feasible trajectory. We propose Physics-Guided Search Distillation (PGSD), a single-view planning framework that combines future representations and execution feedback for verified beyond-pool repair through three components: (1) Execution-Aware Selection checks ego-vehicle road compliance under simulated execution and selects the highest-scoring feasible trajectory; (2) a Clearance-Aware Search Teacher generates beyond-pool corrections and selects demonstrations with larger minimum road-boundary clearance during execution to buffer imitation errors; and (3) a Future-Conditioned Repair Student predicts one correction from rewarder features and execution feedback for single-rollout verification. The pretrained planner and rewarder remain frozen; no future scenes are generated online. On NAVSIM v1, single-view PGSD achieves 90.97 PDMS, exceeding all compared prior single-view planners and retaining 73.2% of the search teacher's PDMS gain over execution-aware selection. On NAVSIM v2 navhard, it reaches 38.96 EPDMS, improving on WorldDrive by 11.24 points (40.5%). The student replaces up to ten search rollouts with a single verification rollout per repair trigger. HUGSIM Hard examples provide qualitative evidence of zero-shot transfer.
Figures & tables
Figure 1: Existing selection and our search distillation. (a) Existing future-aware scoring selects from a fixed candidate pool. (b) Our Clearance-Aware Search Teacher (CAST) generates beyond-pool corrections through repeated execution rollouts. (c) Our Future-Conditioned Repair Student (FCRS) combines world-model (WM) features and execution feedback, predicting one correction checked by Single-Rollout Verification (SRV). Rollout counts denote additional repair queries; dashed warm-gold arrows denote training-only teacher demonstrations.
Figure 2: Overview of PGSD. C denotes candidate trajectories and FAR scores; F , reused future features; and E , reference execution feedback. EAS intervenes when executed penetration reaches 5 cm and selects the highest-scoring candidate with nonnegative clearance when available. The selected reference τEAS conditions repair and serves as fallback. FCRS handles triggered beyond-pool repair; SRV verifies its proposal, with Selection Fallback (SF) on rejection. Snowflakes mark frozen networks; flames mark the trainable future branch. Dashed warm-gold arrows denote offline CAST supervision.
Method
Sensors
World model
NC ↑
DAC ↑
TTC ↑
Comf. ↑
EP ↑
PDMS ↑
Δ PDMS
TransFuser [ 36 ]
MV & L
×
97.7
92.8
92.8
100.0
79.2
84.0
-5.0
Hydra-MDP [ 32 ]
MV & L
×
98.3
96.0
94.6
100.0
78.7
86.5
-2.5
DiffusionDrive [ 3 ]
MV & L
×
98.2
96.2
94.7
100.0
82.2
88.1
-0.9
WoTE [ 6 ]
MV & L
✓
98.5
96.8
94.9
99.9
81.9
88.3
-0.7
UniAD [ 1 ]
MV
×
97.8
91.9
92.9
100.0
78.8
83.4
-5.6
PARA-Drive [ 37 ]
MV
×
97.9
92.4
93.0
99.8
79.3
84.0
-5.0
Table 1: Planning performance on NAVSIM v1 navtest. Baseline scores follow WorldDrive [ 10 ] . SV/MV and L denote single-/multi-view cameras and LiDAR. PGSD also uses road maps. Δ PDMS is the absolute gain over the reported WorldDrive (89.0). Bold/underline mark best/second-best distinct scores, including ties; CAST is a search reference.
Method
Sensors
Stage
NC ↑
DAC ↑
EP ↑
TTC ↑
HC ↑
EPDMS ↑
LTF †
MV
S1
96.2
79.5
84.1
95.1
97.5
23.1
S2
77.7
70.2
85.1
75.6
95.7
DiffusionDrive †
MV
S1
96.8
86.0
84.0
95.8
97.6
27.5
S2
80.1
72.8
85.9
76.6
96.3
WorldDrive
SV
S1
95.7
80.4
81.8
94.4
97.6
27.72
S2
82.9
68.0
81.6
78.2
97.4
Table 2: Planning performance on NAVSIM v2 navhard. S1 and S2 denote the two evaluation stages; EPDMS is the aggregate two-stage score. SV/MV denote single-/multi-view cameras. † Results reported in WorldDrive [ 10 ] . We reproduce the WorldDrive baseline using its candidate-score selection. All scores use the 0–100 scale. Bold/underline indicate the best/second-best distinct scores per stage and aggregate, including ties.
Figure 3: Planning quality and additional repair cost. (a) PDMS versus additional rollouts per trigger; the horizontal line marks EAS. (b) Gain over EAS versus cached repair latency.
Figure 4: Demonstration clearance and execution feedback. (a,b) Demonstration selection with full execution feedback (three paired seeds): (a) training action Huber loss (mean ± sample SD); (b) qualifying rates versus required clearance, averaged over seeds. (c) Effect of execution feedback under road-margin supervision. Panels (b,c) evaluate the same 241 development trajectories.
Figure 5: Execution-aware selection on NAVSIM. Each panel pairs a front-camera view with a bird’s-eye view of four-second simulated execution. Red dashed lines denote WorldDrive; blue solid lines denote execution-aware selection. Panels show (a) obstacle bypassing, (b) a roundabout, (c) rainy conditions, (d) close interaction, (e) turning near a boundary, (f) a narrow bend, (g) rain with obstacles, and (h) turning near traffic. Surrounding actors are shown at the observation time.
Method
EAS
Repair
FAR → student
SRV
DAC ↑
EP ↑
PDMS ↑
Component variants
WorldDrive
—
None
—
—
97.0
83.5
89.0
+ EAS
✓
None
—
—
98.4
84.6
90.4
+ Fixed rule
✓
Rule
—
✓
98.7
84.8
90.6
+ PRS
✓
Student
—
✓
98.8
84.9
90.7
Full framework and teacher reference
Table 3: Component analysis on NAVSIM v1. WorldDrive is reproduced. All variants retain FAR for candidate scoring; FAR → student denotes FAR features used for repair.
Method
PDMS ↑
Δ vs. EAS ↑
Gain retained (%) ↑
Rollouts / trigger ↓
Median latency (ms) ↓
Selection and repair references
EAS
90.4
+0.00
—
0.00
—
Fixed rule
90.6
+0.22
28.7
1.00
23.96
PRS
90.7
+0.32
41.3
1.00
24.85
Future-conditioned repair and search
FCRS
91.0
+0.56
73.2
1.00
25.13
Table 4: Distillation quality and additional repair cost. Gains and retention ratios use the common EAS reference and CAST, computed before rounding. Rollouts are mean additional queries per repair trigger; FCRS uses a fixed budget of one. Latencies measure the cached single-thread CPU correction stage in milliseconds. Dashes denote inapplicable entries.
Appendix figures & tables4 assets
Supplementary material from the paper’s appendix.
Appendix
Figure S1 : FCRS architecture. The future feature of the EAS reference is projected to 32 dimensions and concatenated with the frozen PRS’s 64-dimensional hidden representation. A two-dimensional residual is added to the original repair parameters before deformation and verification. Blue marks frozen paths, warm gold the trainable branch, and purple deterministic steps.
Figure S2 : Origin of the inherited FAR. In WorldDrive, a frozen trajectory-aware world model supplies future latent targets for FAR pretraining. The Future-scene Decoder predicts candidate-conditioned latent features aligned to these targets through MSE supervision. The Future-aware Decoder and reward MLP score candidates using the distilled representation. Dashed warm-gold arrows denote WorldDrive training supervision. PGSD uses the resulting pretrained FAR and keeps it frozen.
Figure S3 : Execution-clearance distributions. Each dot is one reference trajectory’s mean minimum executed clearance across three seeds. Half-densities share a Gaussian-kernel bandwidth and vertical scale. Boxes show median and IQR; whiskers show the 5th and 95th percentiles. The full linear range retains negative tails. Pale gold denotes negative clearance; the dashed line marks 0.02 m.
Figure S6 : Qualitative zero-shot planning examples on HUGSIM Hard. Selected front-camera views without HUGSIM-specific training or fine-tuning show (a) an oncoming encounter, (b) a narrow passage, (c) urban traffic, (d) a nearby vehicle, (e) a curved roadway, and (f) lead-vehicle following. Yellow lines and red dots denote paths and waypoints. Each panel shows one planning snapshot.
Autonomous driving requires choosing a safe and efficient plan as surrounding traffic evolves. Generate-and-select planners propose multiple trajectories and score them for execution, and they have outperformed representative direct-prediction baselines on NAVSIM. Their scorer must compare plans that were never executed. Driving logs record the future of only the executed trajectory, so matching the logged future can leave predictions for the alternatives unconstrained; a simulator, in contrast, can label the outcome of every candidate. We introduce World4Scorer, which builds the scorer as a trajectory-conditioned JEPA-style predictor: it predicts a state for each candidate and reads the candidate's scores from that state. Simulator outcome labels supervise the states of all candidates, and the observed future of the executed trajectory anchors the predictor to real scene evolution. Because one predictor produces every candidate's state, the anchor can constrain shared parameters used to score unexecuted plans, while the future itself is needed only during training. Generated candidates mostly score well, so a scene-matched bank adds low-scoring plans to the outcome supervision; framewise choices can conflict, so inertial re-ranking keeps consecutive selections consistent. World4Scorer achieves state-of-the-art NAVSIM-v2 performance and a strong adapted-system result on closed-loop Bench2Drive. With the LeWM world model and planning budget fixed, outcome-based scoring also improves manipulation planning on the OGBench-Cube benchmark.
Jieyuan Pei, Meiyi Lu, Sining Ang +12
Institute for AI Industry Research (AIR), Tsinghua University · HiThink Research · The Hong Kong University of Science and Technology (Guangzhou) +8
Generalization is a central challenge in autonomous driving, as real-world deployment requires robust performance under unseen scenarios, sensor domains, and environmental conditions. Recent world-model-based planning methods have shown strong capabilities in scene understanding and multi-modal future prediction, yet their generalization across datasets and sensor configurations remains limited. In addition, their loosely coupled planning paradigm often leads to poor video-trajectory consistency during visual imagination. To overcome these limitations, we propose DriveVA, a novel autonomous driving world model that jointly decodes future visual forecasts and action sequences in a shared latent generative process. DriveVA inherits rich priors on motion dynamics and physical plausibility from well-pretrained large-scale video generation models to capture continuous spatiotemporal evolution and causal interaction patterns. To this end, DriveVA employs a DiT-based decoder to jointly predict future action sequences (trajectories) and videos, enabling tighter alignment between planning and scene evolution. We also introduce a video continuation strategy to strengthen long-duration rollout consistency. DriveVA achieves an impressive PDM-based planning performance of 90.9 PDM score on the NAVSIM benchmark. Extensive experiments also demonstrate the zero-shot capability and cross-domain generalization of DriveVA, which reduces average L2 error and collision rate by 78.9% and 83.3% on nuScenes and 52.5% and 52.4% on the Bench2Drive built on CARLA v2 compared with the state-of-the-art world-model-based planner.
Mengmeng Liu, Diankun Zhang, Jiuming Liu +7
University of Twente, The Netherlands · Xiaomi EV, China · University of Cambridge, United Kingdom +1
World models in end-to-end autonomous driving predict future scene evolution to provide foresight for trajectory planning. Existing methods mainly study how to predict the future and how to use it, but less often ask which future representation is actually most useful for planning. To this end, we propose PlanWAM, a Planning-Shaped World Action Model. The key idea is to let the planning task shape the future-state representation, so that it retains the information most useful for planning. A latent world model then predicts this planning-shaped future latent representation from historical observations and uses it for planning, enabling foresighted planning. Specifically, we first use a Temporal Register Pyramid to compress multi-frame historical information in a recency-aware manner, learning a compact history representation oriented toward future reasoning and planning. We then introduce a privileged future posterior branch that observes ground-truth future frames, and shape its future latent representation with trajectory-planning objectives to obtain a planning-shaped future latent representation. Hindsight-to-Foresight Distillation trains a prior branch that depends only on history to predict this future latent representation. The predicted future latent representation serves as planning context and guides trajectory generation and selection. PlanWAM achieves 93.8 PDMS / 90.9 EPDMS on NAVSIM-v1/v2 navtest and reaches 38.7 HD-Score on closed-loop HUGSIM in a zero-shot setting, demonstrating leading planning performance across both open-loop and closed-loop evaluations. Extensive experiments further demonstrate that planning-shaped future representations provide an effective and deployable form of foresight for world-action models.
Jinchang Xu, Hongda Yu, Fengwei Dong +9
Southeast University · NIO · Nanyang Technological University +1