Paper deep dive

Planning-Aligned Pretraining of BEV Representations with Sparse Action-Conditioned Targets for End-to-End Autonomous Driving

Jaeha Song, Soonmin Hwang

Year: 2026Venue: arXiv preprintArea: cs.CVType: PreprintEmbeddings: 95

Intelligence

Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 96%

Last extracted: 9/22/2026, 5:21:36 AM

Summary

The paper introduces PAVER, a planning-aligned pretraining method for Bird's-Eye-View (BEV) representations in end-to-end autonomous driving. PAVER constructs sparse risk and unknown targets from single LiDAR sweeps along rule-based ego motions, using a lightweight 10K-parameter head to predict these targets from masked BEV features. This approach avoids dense reconstruction and task annotations, transferring only the BEV encoder. Experiments on nuScenes and Bench2Drive show significant improvements in planning, collision rates, and driving scores compared to baselines like VAD and UniAD, while reducing total training time.

Entities (10)

BEV · concept · 100%LiDAR · sensor · 100%nuScenes · dataset · 100%PAVER · method · 100%Bench2Drive Town05 Long · dataset · 95%UniAD-Tiny · model · 95%VAD-Tiny · model · 95%Collision Rate · metric · 90%Driving Score · metric · 90%Planning L2 · metric · 90%

Relation Signals (7)

PAVER → evaluatedon → nuScenes

confidence 100% · On nuScenes, PAVER reduces VAD-Tiny's average collision rate

PAVER → uses → LiDAR

confidence 100% · From a single LiDAR sweep, PAVER constructs sparse risk and unknown targets

PAVER → evaluatedon → Bench2Drive Town05 Long

confidence 95% · On Bench2Drive Town05 Long, PAVER improves UniAD-Tiny's closed-loop Driving Score

PAVER → improves → UniAD-Tiny

confidence 95% · On Bench2Drive Town05 Long, PAVER improves UniAD-Tiny's closed-loop Driving Score from 48.45 to 58.79

PAVER → improves → VAD-Tiny

confidence 95% · On nuScenes, PAVER reduces VAD-Tiny's average collision rate from 0.51% to 0.19%

PAVER → increases → Driving Score

confidence 95% · PAVER improves UniAD-Tiny's closed-loop Driving Score from 48.45 to 58.79

PAVER → reduces → Collision Rate

confidence 95% · PAVER reduces VAD-Tiny's average collision rate from 0.51% to 0.19%

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:End-to-end driving requires planning-relevant bird's-eye-view (BEV) representations, but existing pretraining approaches often rely on task annotations or dense scene reconstruction. We introduce PAVER, Planning-Aligned BEV Encoder Pretraining. From a single LiDAR sweep, PAVER constructs sparse risk and unknown targets describing occupied and unobserved evidence along rule-based ego motions. A 10K-parameter head predicts these targets from masked BEV features conditioned on the action state, directing supervision toward geometric constraints on candidate motions. Pretraining requires no driving-task annotations or dense reconstruction. Only the BEV encoder is transferred, preserving the downstream architecture and camera-only inference. On nuScenes, PAVER reduces VAD-Tiny's average collision rate from 0.51% to 0.19%, while improving planning L2, motion prediction, detection, and mapping. The selected VAD-Tiny and VAD-Base schedules use about 36% less estimated total training time than scratch training, including pretraining. On Bench2Drive Town05 Long, PAVER improves UniAD-Tiny's closed-loop Driving Score from 48.45 to 58.79. The project page is available at this https URL.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

91,408 characters extracted from source content.

Expand or collapse full text

Planning-Aligned Pretraining of BEV Representations with Sparse Action-Conditioned Targets for End-to-End Autonomous Driving Jaeha Song, Soonmin Hwang ∗ Hanyang University archiiive99@hanyang.ac.kr, soonminh@hanyang.ac.kr Abstract End-to-end driving requires planning-relevant bird’s-eye- view (BEV) representations, but existing pretraining ap- proaches often rely on task annotations or dense scene re- construction. We introduce PAVER, Planning-Aligned BEV Encoder Pretraining. From a single LiDAR sweep, PAVER constructs sparse risk and unknown targets describing occu- pied and unobserved evidence along rule-based ego motions. A 10K-parameter head predicts these targets from masked BEV features conditioned on the action state, directing su- pervision toward geometric constraints on candidate motions. Pretraining requires no driving-task annotations or dense re- construction. Only the BEV encoder is transferred, preserving the downstream architecture and camera-only inference. On nuScenes, PAVER reduces VAD-Tiny’s average collision rate from 0.51% to 0.19%, while improving planning L2, motion prediction, detection, and mapping. The selected VAD-Tiny and VAD-Base schedules use about 36% less estimated total training time than scratch training, including pretraining. On Bench2Drive Town05 Long, PAVER improves UniAD-Tiny’s closed-loop Driving Score from 48.45 to 58.79. The project page is available at https://archiiive99.github.io/PAVER. Introduction Camera-based end-to-end driving combines detection, map prediction, motion prediction, and planning through a shared bird’s-eye-view (BEV) representation (Hu et al. 2022, 2023; Jiang et al. 2023; Sun et al. 2024; Zheng et al. 2024; Liao et al. 2025). Learning cross-view geometry and planning-relevant features requires long downstream training schedules (Hu et al. 2023; Jiang et al. 2023). Image pretraining initializes the image backbone, but BEV construction must still be learned downstream (Deng et al. 2009; He et al. 2016; Li et al. 2022). Many systems first train perception, tracking, or mapping before fine-tuning the full end-to-end model. This approach depends on task-specific annotations, heads, and decoder de- sign (Hu et al. 2022, 2023; Jiang et al. 2023; Weng et al. 2024; Doll et al. 2024; Tang et al. 2025; Li and Cui 2025). PAVER instead pretrains BEV features at locations queried by candidate ego motions. Driving-specific pretraining methods provide relevant ge- ometry, yet representative objectives remain centered on re- constructing scene content (Yang et al. 2024a; Zou et al. ∗ Corresponding author. Figure 1: PAVER improves planning L2 on VAD-Tiny and VAD-Base with about 36% less estimated total training time, including pretraining (Table 9). The lower panels compare planning performance and detection NDS across pretraining methods and auxiliary-head sizes. 2025; Zhang et al. 2025). As summarized in Fig. 2, these methods supervise general spatial support rather than the re- gions most relevant to planning. Planning instead depends on which observed regions constrain candidate actions (Hu et al. 2023). This mismatch motivates supervising the BEV regions used for planning. We introduce PAVER, Planning-Aligned BEV Encoder Pretraining, which pretrains an otherwise unchanged BEV encoder. Following occupancy-grid mapping (Elfes 1989), a LiDAR sweep is reduced to sparse free, occupied, and un- known evidence. Rule-based actions index supervised BEV regions, where samples across the vehicle footprint form sparse action targets for risk and unknown evidence. PAVER replaces these corridor features with a shared mask token and predicts the corresponding sparse action targets conditioned on the action state. This directs BEV pretraining toward evi- dence that constrains candidate actions. arXiv:2609.22868v1 [cs.CV] 19 Sep 2026 Figure 2: Driving pretraining paradigms. Existing work su- pervises pretraining with (a) perception labels or (b) dense reconstruction; (c) PAVER introduces sparse action targets. PAVER requires neither task-specific annotations, a learned teacher, nor dense 3D scene reconstruction. The pretraining-only PAVER head adds 10K parameters and is removed before downstream training; only the BEV encoder is transferred, preserving the downstream architecture and inference pipeline. With this lightweight supervision and shorter total training time (Figure 1), PAVER matches or im- proves the corresponding baseline on most evaluated multi- task metrics (Tables 9 and 1). Collision behavior nevertheless varies across architectures (Table 4). Our contributions are: • We formulate planning-aligned BEV encoder pretraining with supervision at locations along candidate ego mo- tions. • We construct sparse action targets from LiDAR risk and unknown evidence and predict them with a masked, action-conditioned BEV branch, without scene recon- struction, future prediction, or a learned teacher. • We transfer only the BEV encoder and discard the 10K- parameter auxiliary head. We evaluate multi-task transfer and planning performance across several architectures. Related Work Shared representations and initialization in end-to-end driving. End-to-end driving systems couple perception, prediction, and planning through a shared spatial represen- tation, ranging from dense BEV grids (Li et al. 2022, 2023; Huang et al. 2023) to vectorized or sparse scene elements (Hu et al. 2023; Jiang et al. 2023; Sun et al. 2024; Zheng et al. 2024; Li and Cui 2025). Initialization is commonly obtained from downstream supervision: ST-P3, UniAD, and VAD stage perception-oriented training before the broader task stack, PARA-Drive redesigns task connectivity for joint optimization, and later systems vary the decoder itself (Hu et al. 2022, 2023; Jiang et al. 2023; Weng et al. 2024; Doll et al. 2024; Tang et al. 2025; Li and Cui 2025). The repre- sentation is therefore shaped by model-specific annotations and decoders. PAVER separates initialization from that opti- mization, targeting architectures that retain a dense, spatially indexed BEV representation and leaving their task heads un- changed. Driving-specific 3D pretraining. Reconstruction methods learn transferable geometry from masked or projected scene content. UniPAD reconstructs continuous 3D structure and appearance, MIM4D extends reconstruction to multi-view video, and VisionPAD uses rendering and temporal consis- tency to learn 3D-aware features (Yang et al. 2024a; Zou et al. 2025; Zhang et al. 2025). These methods lift image features into a 3D volume and supervise it densely, so the pretrained weights that transfer are the image backbone rather than the shared BEV encoder. BEV-MAE and Occupancy-MAE sim- ilarly reconstruct masked spatial structure for LiDAR repre- sentations (Lin et al. 2024; Min et al. 2022). A separate line distills LiDAR knowledge into camera models through fea- ture or label alignment (Chen et al. 2023; Wang et al. 2023; Zhou et al. 2023; Kim et al. 2024; Hinton, Vinyals, and Dean 2015; Vapnik and Izmailov 2015). PAVER pretrains the BEV encoder itself, uses no learned LiDAR encoder, and performs no feature matching: deterministic free, occu- pied, and unknown evidence is aggregated into sparse action targets. Predictive world models. World-model pretraining learns representations by predicting how scenes evolve. DriveWorld models spatiotemporal latent dynamics, while ViDAR pre- dicts future geometric observations from camera history (Min et al. 2024; Yang et al. 2024b). LAW, DriveX, and DLWM connect future scene or latent prediction to plan- ning (Li et al. 2025; Shi et al. 2025; Zhu et al. 2026). These objectives use future observations, geometry, or learned tran- sition targets. PAVER instead supervises current geometric evidence from a single synchronized LiDAR sweep. Figure 3: Two-stage training protocol. Only the pretrained BEV encoder is transferred, and the existing end-to-end model is fine-tuned unchanged. Figure 4: PAVER overview. PAVER constructs sparse action targets for risk and unknown evidence, masks their associated BEV regions, and predicts the targets from the remaining context. The PAVER head is removed before downstream training. Action and planning signals in representation learning. Actions have appeared in driving pretraining as labels, con- ditioning variables, policy targets, or guides for latent predic- tion (Zhang, Peng, and Zhou 2022; Xiao et al. 2021; Wang et al. 2026; Chen et al. 2020; Li et al. 2025; Zhu et al. 2026). PAVER instead uses rule-based actions to index current Li- DAR evidence, not as behavioral prediction targets. Method Overview PAVER pretrains the BEV encoder with sparse action tar- gets constructed from LiDAR evidence along rule-based ego motions. Multi-view images I are processed by the same BEV encoder used for downstream end-to-end training (Lin et al. 2017; Li et al. 2022; Jiang et al. 2023), producing F = f bev (I) ∈ R B×C×H×W . LiDAR measurements are used only to construct pretraining targets. Detection, map- ping, motion-prediction, and planning annotations are not used, and the corresponding task decoders are disabled dur- ing pretraining. All parameters of f bev are optimized by the pretraining objective and transferred to downstream training (Fig. 3). The action generator, LiDAR target builder, and the PAVER head (Fig. 3), which comprises the learnable mask token, pointwise projector, and prediction MLP, are used only during pretraining. Candidate Ego Motions Rule-based candidate motions define the supervision loca- tions; they are not expert trajectories or learned action pro- posals. Finite sets of longitudinal accelerations and yaw rates are combined into rule-based control families, tiled across action slots, and independently perturbed during training. Each con- trol is rolled out from the current ego speed over T horizons, giving the ego-frame state q ego s,t = (x ego s,t ,y ego s,t ,ψ ego s,t ,v s,t ) for action slot s and horizon t, where ψ is the yaw heading. Rollout positions and headings are transformed into the LiDAR/BEV indexing frame, yielding ̄q bev s,t = ( ̄x s,t , ̄y s,t , ̄ ψ s,t ,v s,t ). The transformed state is used for Li- DAR target construction and BEV feature sampling, while the PAVER head is conditioned on the original ego-frame state. Figure 5: Sparse action-target construction. LiDAR rays pro- vide Risk and Unknown targets for action-indexed queries; “Safe” denotes Free evidence. Sparse LiDAR Evidence We rasterize LiDAR measurements into a BEV evidence map E ∈ Unknown, Free, Occupied B×H×W following occupancy-grid mapping (Elfes 1989) (Figure 5). Cells start as Unknown; samples along each open ray segment are marked Free (Amanatides and Woo 1987), and measured endpoints override their cells as Occupied. This map is used only to construct sparse targets; the camera model does not predict dense occupancy. Sparse Action Targets At each candidate state ̄q bev s,t we query K locations laterally across the vehicle width, andV s,t denotes the queries that stay in range. The two targets are the fractions of those queries that fall on measured endpoints and on cells without supporting ray evidence: r s,t = |k ∈V s,t : E k = Occupied| |V s,t | ,(1) EpochsAuxiliary TasksPlanningMotionDetectionMap MethodPTFT PretrainingFine-tuningL2 (m)↓ Col. (%)↓ ADE↓ FDE↓ MR↓ mAP↑ NDS↑ mAP↑ VAD-Tiny † 060 NoneDet., Map, Mot., Plan. 0.660.510.91 1.25 0.13 0.23 0.34 0.42 + Ours2030ActionsDet., Map, Mot., Plan.0.600.190.801.090.120.280.400.44 VAD-Base † 060 NoneDet., Map, Mot., Plan. 0.740.310.76 1.03 0.11 0.29 0.42 0.50 + Ours2030ActionsDet., Map, Mot., Plan.0.560.400.690.900.090.330.450.50 GenAD † 060 NoneDet., Map, Mot., Plan. 0.590.370.87 1.18 0.14 0.19 0.26 0.46 + Ours2030ActionsDet., Map, Mot., Plan.0.540.210.801.050.110.210.280.44 † Reproduced results. Table 1: Multi-task performance on the nuScenes validation set. Actions denotes PAVER’s sparse action targets. PT and FT denote pretraining and fine-tuning epochs; Det., Mot., and Plan. denote detection, motion prediction, and planning. L2 is the trajectory displacement error and Col. the box collision rate in percent, both averaged over 1, 2, and 3 seconds; ADE, FDE, and MR are the motion average and final displacement errors and miss rate; mAP and NDS are mean average precision and the nuScenes detection score. u s,t = |k ∈V s,t : E k = Unknown| |V s,t | ,(2) where E k is the evidence value at the k-th query position. Risk measures observed occupancy, while unknown mea- sures missing observations. These continuous ratios describe current evidence, not future collision probabilities or cali- brated uncertainty. Candidate motions index the supervision locations; they are not prediction targets. This planning- aligned supervision places targets along rule-based ego mo- tions rather than densely across the BEV raster. Figure 6: Action-corridor feature masking. Green markers indicate lateral-query locations. Selected BEV cells are re- placed by a shared learnable token; other cells are unchanged. Action-Corridor Feature Masking To limit direct access to target-associated features, we form a binary mask M ∈ 0, 1 B×1×H×W from the union of rasterized lateral-query cells. Selected cells in a temporary feature map are replaced by a shared learnable token m ∈ R 1×C×1×1 : F mask = (1− M )⊙ F + M ⊙ m.(3) Unselected cells retain their original features and the original representation F is left unchanged. Figure 6 illustrates the feature replacement. Masking removes the selected local feature vectors, but not information encoded before masking, and does not guarantee the use of neighboring cells. Action-Conditioned Prediction A pointwise 1 × 1 projector φ produces Z = φ(F mask ) without spatial aggregation. For each action slot and horizon, we bilinearly sample z s,t at the transformed action center ( ̄x s,t , ̄y s,t ) (Jaderberg et al. 2015), concatenate the ego-frame action state, and predict two logits with the MLP g: (ˆr s,t , ˆu s,t ) = g z s,t ;q ego s,t .(4) The action-center cell is masked because the lateral-offset set includes zero. The head predicts one risk/unknown pair from a center feature, while each target pair aggregates K lateral LiDAR queries. Pretraining Objective The pretraining objective is L pre = λ r L r + λ u L u ,(5) whereL r andL u are logit-based binary cross-entropy losses for the continuous risk and unknown targets, averaged over valid action–horizon entries. Figure 4 summarizes pretraining. Only the pretrained BEV encoder is transferred; the PAVER head and target-generation components are discarded. The downstream model, training objectives, and camera-only inference are unchanged. The Technical Supplement provides rollout equations, co- ordinate conventions, target-construction details, and train- ing settings. Experiments Implementation Details We pretrain on the 28,130 nuScenes training keyframes and evaluate on all 6,019 validation keyframes, using VAD-Tiny (Jiang et al. 2023) unless stated otherwise. PAVER uses a single-frame temporal queue and LiDAR only for target con- struction. We pretrain for 20 epochs on 4 GPUs with AdamW (Loshchilov and Hutter 2019), cosine annealing (Loshchilov and Hutter 2017), and a reduced image-backbone learn- ing rate. Downstream training follows each model’s original recipe, with the transferred BEV encoder and reinitialized task decoders. All metrics in each row come from a sin- gle checkpoint selected for joint multi-task balance within its training budget. Additional settings are in the Technical Supplement. We evaluate the transferred models on nuScenes (Caesar et al. 2020) using the planning, motion-prediction, detection, and vector-map metrics adopted by VAD (Jiang et al. 2023). Table 1 averages planning over 1, 2, and 3 seconds; per- horizon values are reported separately. PlanningMotionDetection Map Target Source L2 (m)↓ Col. (%)↓ ADE↓ FDE↓ MR↓ NDS↑ mAP↑ None † 0.6620.5130.905 1.250 0.135 0.338 0.419 Pseudo-LiDAR 0.5630.5600.831 1.101 0.121 0.386 0.439 LiDAR0.6030.1870.8031.0860.1210.3970.439 † Reproduced results. Table 2: Planning results with pseudo-LiDAR pretraining. Planning and motion metrics are averaged over 1, 2, and 3 seconds. None denotes the VAD-Tiny baseline. Results with Pseudo-LiDAR When LiDAR is unavailable, PAVER constructs targets by unprojecting per-camera depth (Wang et al. 2019) from frozen UniK3D (Piccinelli et al. 2025). Pseudo-LiDAR im- proves planning L2, motion, detection, and mapping over the baseline (Table 2). Its L2 is lower than with mea- sured LiDAR, but collision rate exceeds both the baseline and measured-LiDAR variants. Estimated geometry there- fore supports multi-task transfer without reproducing mea- sured LiDAR’s collision reduction. Depth-quality analyses are provided in the Technical Supplement. Quantitative Results on nuScenes PAVER matches or improves most downstream metrics on three architectures without task annotations during pretrain- ing (Table 1). On VAD-Tiny and GenAD, detection gains are larger for pedestrians, motorcycles, and bicycles than for cars (Table 3). AP↑ MethodCar Truck Bus Ped. Motor. Bike mAP VAD-Tiny † 0.49 0.19 0.28 0.28 0.13 0.17 0.23 + Ours0.510.240.410.350.220.230.28 VAD-Base † 0.55 0.24 0.33 0.38 0.24 0.27 0.29 + Ours0.550.240.460.420.310.290.33 GenAD † 0.44 – 0.29 0.17 0.22 0.19 + Ours0.46–0.340.230.250.21 † Reproduced results. Table 3: Detection results per object class on nuScenes. Ped., Motor. and Bike denote pedestrian, motorcycle and bicycle. A dash marks a class the reproduced GenAD never detects; the mean still counts it as zero. L2 (m)↓Col. (%)↓ Method1s 2s 3s 1s 2s 3s VAD-Tiny † 0.35 0.63 1.01 0.38 0.45 0.71 + Ours0.320.580.920.090.150.32 VAD-Base † 0.41 0.71 1.10 0.20 0.29 0.43 + Ours0.300.540.850.240.410.55 GenAD † 0.33 0.56 0.88 0.21 0.34 0.56 + Ours0.280.510.820.090.180.35 † Reproduced results. Table 4: Planning results by horizon on nuScenes, reported at the 1, 2, and 3 second horizons. On VAD-Tiny, collision rate falls from 0.51 to 0.19%, planning L2 from 0.66 to 0.60 m, and motion ADE from 0.91 to 0.80 m, while map mAP rises from 0.42 to 0.44. GenAD reduces collision from 0.37 to 0.21% and L2 by 0.05 m. VAD-Base reduces L2 by 0.18 m but increases collision from 0.31 to 0.40%. VAD-Tiny and GenAD improve L2 and collision at ev- ery horizon, whereas VAD-Base improves L2 but increases collision (Table 4). Compared with reconstruction-based pretraining, PAVER achieves lower L2 at every VAD-Tiny horizon with 0.01M auxiliary parameters, compared with 6.41M and 13.43M (Ta- ble 6). Grad-CAM highlights vehicles and pedestrians (Fig- ure 8), while Figure 7 compares downstream predictions. Additional results are in the Technical Supplement. MethodDS↑ RC↑ IS↑ UniAD-Tiny 48.45 60.96 0.85 + Ours58.7979.060.79 GenAD ‡ 34.53– + Ours49.0758.950.90 ‡ Reported by GenAD. Table 5: Closed-loop results on Bench2Drive Town05 Long. DS, RC, and IS denote Driving Score, Route Completion, and Infraction Score, respectively; a dash marks a metric the corresponding source does not report. Closed-loop qualita- tive results are in the Technical Supplement. Closed-Loop Evaluation We evaluate closed-loop transfer on nine Bench2Drive Town05 Long routes (Jia et al. 2024), with one repetition per route (Table 5). On UniAD-Tiny, Driving Score rises from 48.45 to 58.79 and Route Completion from 60.96 to 79.06, while Infraction Score decreases from 0.85 to 0.79. At the route level, PAVER initialization converts three UniAD- Tiny timeouts into completions. Detailed per-route outcomes, completion timelines, and frame-by-frame qualitative com- parisons are provided in the Technical Supplement. Input ImagesGround TruthVAD-Tiny+ PAVER Figure 7: VAD-Tiny qualitative comparisons on nuScenes. More results are in the Technical Supplement. L2 (m)↓Col. (%)↓ MethodTarget#Params 1s 2s 3s 1s 2s 3s VAD-Tiny † –0.00M 0.35 0.63 1.01 0.38 0.45 0.71 + UniPAD ‡ RGB, Depth6.41M 0.45 0.75 1.11 0.19 0.31 0.46 + MIM4D ‡ RGB, Depth13.43M 0.39 0.69 1.06 0.21 0.28 0.38 + OursSparse Actions0.01M0.320.580.920.090.150.32 VAD-Base † –0.00M 0.41 0.71 1.10 0.20 0.29 0.43 + MIM4D ‡ RGB, Depth13.43M 0.36 0.64 1.00 0.08 0.14 0.36 + OursSparse Actions0.01M0.300.540.850.240.410.55 † Reproduced results. ‡ Reported results. Table 6: Planning performance under different pretraining methods. RGB denotes camera-color reconstruction, and Depth denotes LiDAR-projected sparse metric-depth recon- struction. Sparse Actions denotes PAVER’s sparse action tar- gets. #Params reports temporary auxiliary parameters. Ablation Study Comparison of Pretraining Supervision Detection-and-map pretraining gives the lowest planning L2, while PAVER achieves lower collision and higher NDS than the task-supervised alternatives (Table 7). The 0.03M vari- ant widens the projector and action-conditioned MLP; the 0.09M variant replaces the pointwise projector with a 3× 3 projector. Larger heads improve some metrics but increase collision rate. The Technical Supplement compares down- stream learning curves and multi-task transfer across pre- training objectives. Feature Masking and Action State Variants differ only in the ablated component: dropping the concatenated [x,y,ψ,v] input, or sampling the original BEV features at the same action centers. Masking or action conditioning alone gives collision rates of 0.523% and 0.550%, compared with 0.513% for the base- line. Combining them reduces collision to 0.187% and gives PretrainingPlan.Mot. Det. Map Det. Map Occ. † Act. #Params↓ L2 (m)↓ Col. (%)↓ ADE↓ NDS↑ mAP↑ 0.00M 0.6620.513 0.905 0.338 0.419 ✓3.01M 0.5150.240 0.7980.365 0.406 ✓2.91M 0.6940.333 0.857 0.331 0.391 ✓3.56M 0.6460.320 0.905 0.336 0.393 ✓5.92M 0.5040.190 0.805 0.361 0.397 ✓9.49M 0.5770.270 0.7980.356 0.420 ✓0.01M0.6030.1870.8030.3970.439 ✓0.03M0.5140.3200.8000.3830.447 ✓0.09M0.5960.2200.7820.3970.433 † Independently implemented. Table 7: Pretraining strategy ablation. All pretraining runs use 20 epochs. L2 and collision rate are averaged over 1, 2, and 3 seconds. #Params denotes the number of auxiliary parameters. Occ. and Act. denote occupancy and PAVER’s action conditioning. Bold and underline denote the best and second-best results, respectively, among the pretraining runs. the lowest L2 and highest NDS among these variants (Ta- ble 8). Motion and map metrics favor the partial variants. Figure 8 shows image-space attributions. The Technical Supplement provides analyses of target coverage, frozen probes by BEV region, and layer-wise representation sim- ilarity before and after pretraining. Pretraining and Fine-Tuning Schedules Twenty pretraining epochs followed by 30 fine-tuning epochs reduce estimated total training time by about 36%: from 21.3 to 13.6 hours for VAD-Tiny and from 103 to 66.1 hours for VAD-Base, with the lowest L2 in each group (Table 9). VAD- Base’s collision rate is lowest with 20 fine-tuning epochs, while its scratch baseline retains the highest map mAP. PlanningMotionDetection Map Mask Action L2 (m)↓ Col. (%)↓ ADE↓ FDE↓ MR↓ NDS↑ mAP↑ 0.6620.513 0.905 1.250 0.135 0.338 0.419 ✓0.6550.523 0.808 1.084 0.115 0.396 0.443 ✓0.7050.550 0.797 1.080 0.123 0.388 0.445 ✓0.6030.1870.8031.0860.1210.3970.439 Table 8: Component ablation. Mask denotes action-corridor feature masking with a shared learnable token, and Action denotes conditioning the prediction head on the ego-frame action state [x,y,ψ,v], where ψ is the yaw heading. The variant using both is PAVER; the empty row is the VAD- Tiny baseline. Input ImageVAD-Tiny + PAVER Figure 8: Grad-CAM for the BEV target on nuScenes. With PAVER initialization, activations highlight scene objects, in- cluding vehicles and pedestrians. More results are provided in the Technical Supplement. Discussion and Limitations Multi-task performance. PAVER improves planning L2, motion prediction, and detection on all three nuScenes back- bones; VAD-Tiny also improves mapping (Table 1). On VAD- Tiny, motion ADE decreases from 0.91 to 0.80 m and detec- tion NDS increases from 0.34 to 0.40. Sparse action targets can benefit perception and prediction without task-specific pretraining labels. Gains still vary by metric: VAD-Base’s collision rate increases, while GenAD’s map mAP decreases. Closed-loop Driving Score improves for both UniAD-Tiny and GenAD, but UniAD-Tiny’s Infraction Score decreases. Explainability. Object returns contribute occupied evi- dence along candidate motions, providing a geometric prior without object-category labels. VAD-Tiny’s map mAP in- creases from 0.419 to 0.439 despite the absence of map targets during pretraining (Table 9). Grad-CAM highlights vehicles and pedestrians (Figure 8). Frozen-probe analyses are provided in the Technical Supplement. Limitations. PAVER’s targets describe current geometry, not future action outcomes. All horizons query the same sweep, so the targets do not capture obstacles moving into or out of a candidate path. They also encode occupancy and observability rather than traffic rules or right of way; an observed-free corridor need not be a valid driving action. Finally, the rule-based motion sampler leaves unsampled re- gions without direct target supervision during pretraining. Training CostPlanningMotion Detection Map MethodPT FT Time L2 (m)↓ Col. (%)↓ ADE↓ NDS↑ mAP↑ VAD-Tiny † 48 ‡ 12 20.5h 0.7010.730 0.840 0.363 0.413 VAD-Tiny † 0 60 21.3h 0.6620.513 0.905 0.338 0.419 + Ours20106.5h0.6370.3000.8510.3570.361 + Ours202010.1h0.6470.2200.7810.3870.425 + Ours203013.6h0.6030.1870.8030.3970.439 VAD-Base † 0 60 103h 0.7420.3070.760 0.422 0.502 + Ours201031.8h0.6460.5300.7720.3710.403 + Ours202049.0h0.7570.2800.6980.4420.460 + Ours203066.1h0.5620.4010.6850.4460.498 † Reproduced results. ‡ Stage 1 pretrains detection, mapping, and motion prediction without planning. Table 9: Training schedule ablation. Total training time, in- cluding pretraining, is estimated from mean iteration times on 4 NVIDIA RTX 5090 GPUs. Bold marks the best result and underline marks the second-best result. Conclusion PAVER shows that planning-relevant BEV representations can be learned through a simple, interpretable pretraining task without dense scene reconstruction or semantic-task supervision. Predicting risk and unknown evidence along candidate ego motions gives the objective a clear geomet- ric meaning: observed obstacles and missing evidence. No driving-task annotations are required. The compact predic- tion task keeps auxiliary computation small, while the down- stream gains under shorter total training schedules suggest that this initialization can facilitate subsequent optimization. Its benefits extend beyond planning to perception and predic- tion through the shared BEV encoder. These results highlight a route to efficient end-to-end driving: simplify the auxiliary task while retaining relevant geometric evidence. Such su- pervision can improve multi-task performance with reduced training time and without semantic annotation costs during pretraining, while preserving the downstream architecture and camera-only inference. References Amanatides, J.; and Woo, A. 1987. A Fast Voxel Traversal Algorithm for Ray Tracing. In Eurographics, 3–10. Caesar, H.; Bankiti, V.; Lang, A. H.; Vora, S.; Liong, V. E.; Xu, Q.; Krishnan, A.; Pan, Y.; Baldan, G.; and Beijbom, O. 2020. nuScenes: A Multimodal Dataset for Autonomous Driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 11621–11631. Chen, D.; Zhou, B.; Koltun, V.; and Krähenbühl, P. 2020. Learning by Cheating. In Proceedings of the Conference on Robot Learning, volume 100 of Proceedings of Machine Learning Research, 66–75. PMLR. Chen, Z.; Li, Z.; Zhang, S.; Fang, L.; Jiang, Q.; and Zhao, F. 2023. BEVDistill: Cross-Modal BEV Distillation for Multi- View 3D Object Detection. In International Conference on Learning Representations. Deng, J.; Dong, W.; Socher, R.; Li, L.-J.; Li, K.; and Fei- Fei, L. 2009. ImageNet: A Large-Scale Hierarchical Image Database. In Proceedings of the IEEE Conference on Com- puter Vision and Pattern Recognition, 248–255. Devlin, J.; Chang, M.-W.; Lee, K.; and Toutanova, K. 2019. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In Proceedings of the 2019 Con- ference of the North American Chapter of the Association for Computational Linguistics, 4171–4186. Doll, S.; Hanselmann, N.; Schneider, L.; Schulz, R.; Cordts, M.; Enzweiler, M.; and Lensch, H. P. A. 2024. DualAD: Disentangling the Dynamic and Static World for End-to-End Driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 14728–14737. Elfes, A. 1989. Using Occupancy Grids for Mobile Robot Perception and Navigation. Computer, 22(6): 46–57. Fawcett, T. 2006. An Introduction to ROC Analysis. Pattern Recognition Letters, 27(8): 861–874. He, K.; Zhang, X.; Ren, S.; and Sun, J. 2016. Deep Residual Learning for Image Recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 770–778. Hinton, G.; Vinyals, O.; and Dean, J. 2015. Distill- ing the Knowledge in a Neural Network. arXiv preprint arXiv:1503.02531. Hu, S.; Chen, L.; Wu, P.; Li, H.; Yan, J.; and Tao, D. 2022. ST-P3: End-to-End Vision-Based Autonomous Driving via Spatial-Temporal Feature Learning. In European Conference on Computer Vision, 533–549. Hu, Y.; Yang, J.; Chen, L.; Li, K.; Sima, C.; Zhu, X.; Chai, S.; Du, S.; Lin, T.; Wang, W.; Lu, L.; Jia, X.; Liu, Q.; Dai, J.; Qiao, Y.; and Li, H. 2023. Planning-Oriented Autonomous Driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 17853–17862. Huang, Y.; Zheng, W.; Zhang, Y.; Zhou, J.; and Lu, J. 2023. Tri-Perspective View for Vision-Based 3D Semantic Occu- pancy Prediction. In Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition, 9223– 9232. Jaderberg, M.; Simonyan, K.; Zisserman, A.; and Kavukcu- oglu, K. 2015. Spatial Transformer Networks. In Advances in Neural Information Processing Systems, volume 28. Jia, X.; Yang, Z.; Li, Q.; Zhang, Z.; and Yan, J. 2024. Bench2Drive: Towards Multi-Ability Benchmarking of Closed-Loop End-to-End Autonomous Driving. In Advances in Neural Information Processing Systems, Datasets and Benchmarks Track. Jiang, B.; Chen, S.; Xu, Q.; Liao, B.; Chen, J.; Zhou, H.; Zhang, Q.; Liu, W.; Huang, C.; and Wang, X. 2023. VAD: Vectorized Scene Representation for Efficient Autonomous Driving. In Proceedings of the IEEE/CVF International Con- ference on Computer Vision, 8340–8350. Kim, S.; Kim, Y.; Hwang, S.; Jeong, H.; and Kum, D. 2024. LabelDistill: Label-Guided Cross-Modal Knowledge Distil- lation for Camera-Based 3D Object Detection. arXiv preprint arXiv:2407.10164. Li, P.; and Cui, D. 2025. Navigation-Guided Sparse Scene Representation for End-to-End Autonomous Driving. In In- ternational Conference on Learning Representations. Li, Y.; Fan, L.; He, J.; Wang, Y.; Chen, Y.; Zhang, Z.; and Tan, T. 2025. Enhancing End-to-End Autonomous Driving with Latent World Model. In International Conference on Learning Representations. Li, Y.; Ge, Z.; Yu, G.; Yang, J.; Wang, Z.; Shi, Y.; Sun, J.; and Li, Z. 2023. BEVDepth: Acquisition of Reliable Depth for Multi-View 3D Object Detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, 1477–1485. Li, Z.; Wang, W.; Li, H.; Xie, E.; Sima, C.; Lu, T.; Qiao, Y.; and Dai, J. 2022. BEVFormer: Learning Bird’s-Eye-View Representation from Multi-Camera Images via Spatiotem- poral Transformers. In European Conference on Computer Vision, 1–18. Liao, B.; Chen, S.; Yin, H.; Jiang, B.; Wang, C.; Yan, S.; Zhang, X.; Li, X.; Zhang, Y.; Zhang, Q.; and Wang, X. 2025. DiffusionDrive: Truncated Diffusion Model for End-to-End Autonomous Driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 12037–12047. Lin, T.-Y.; Dollár, P.; Girshick, R.; He, K.; Hariharan, B.; and Belongie, S. 2017. Feature Pyramid Networks for Ob- ject Detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2117–2125. Lin, Z.; Wang, Y.; Qi, S.; Dong, N.; and Yang, M.-H. 2024. BEV-MAE: Bird’s Eye View Masked Autoencoders for Point Cloud Pre-training in Autonomous Driving Scenarios. In Proceedings of the AAAI Conference on Artificial Intelli- gence, volume 38, 3531–3539. Loshchilov, I.; and Hutter, F. 2017. SGDR: Stochastic Gradi- ent Descent with Warm Restarts. In International Conference on Learning Representations. Loshchilov, I.; and Hutter, F. 2019. Decoupled Weight Decay Regularization. In International Conference on Learning Representations. Min, C.; Xu, X.; Zhao, D.; Xiao, L.; Nie, Y.; and Dai, B. 2022. Self-Supervised Pre-Training Large-Scale LiDAR Point Clouds with Masked Occupancy Autoencoders. arXiv preprint arXiv:2206.09900. Min, C.; Zhao, D.; Xiao, L.; Zhao, J.; Xu, X.; Zhu, Z.; Jin, L.; Li, J.; Guo, Y.; Xing, J.; Jing, L.; Nie, Y.; and Dai, B. 2024. DriveWorld: 4D Pre-trained Scene Understanding via World Models for Autonomous Driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 15522–15533. Piccinelli, L.; Sakaridis, C.; Segu, M.; Yang, Y.-H.; Li, S.; Abbeloos, W.; and Van Gool, L. 2025. UniK3D: Uni- versal Camera Monocular 3D Estimation. arXiv preprint arXiv:2503.16591. Shi, C.; Shi, S.; Sheng, K.; Zhang, B.; and Jiang, L. 2025. DriveX: Omni Scene Modeling for Learning Generalizable World Knowledge in Autonomous Driving. In Proceedings of the IEEE/CVF International Conference on Computer Vi- sion, 28599–28609. Sun, W.; Lin, X.; Shi, Y.; Zhang, C.; Wu, H.; and Zheng, S. 2024. SparseDrive: End-to-End Autonomous Driv- ing via Sparse Scene Representation. arXiv preprint arXiv:2405.19620. Tang, Y.; Xu, Z.; Meng, Z.; and Cheng, E. 2025. HiP-AD: Hi- erarchical and Multi-Granularity Planning with Deformable Attention for Autonomous Driving in a Single Decoder. In Proceedings of the IEEE/CVF International Conference on Computer Vision. Vapnik, V.; and Izmailov, R. 2015. Learning Using Privi- leged Information: Similarity Control and Knowledge Trans- fer. Journal of Machine Learning Research, 16(1): 2023– 2049. Wang, L.; Yang, Z.; Bai, C.; Zhang, G.; Liu, X.; Zheng, X.; Long, X.-X.; Lu, C.-T.; and Lu, C. 2026. Drive-JEPA: Video JEPA Meets Multimodal Trajectory Distillation for End-to- End Driving. arXiv preprint arXiv:2601.22032. Wang, Y.; Chao, W.-L.; Garg, D.; Hariharan, B.; Campbell, M.; and Weinberger, K. Q. 2019. Pseudo-LiDAR from Visual Depth Estimation: Bridging the Gap in 3D Object Detection for Autonomous Driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 8445–8453. Wang, Z.; Li, D.; Luo, C.; Xie, C.; and Yang, X. 2023. Dis- tillBEV: Boosting Multi-Camera 3D Object Detection with Cross-Modal Knowledge Distillation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 8637–8646. Weng, X.; Ivanovic, B.; Wang, Y.; Wang, Y.; and Pavone, M. 2024. PARA-Drive: Parallelized Architecture for Real- time Autonomous Driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 15449–15458. Xiao, Y.; Codevilla, F.; Pal, C.; and Lopez, A. M. 2021. Action-Based Representation Learning for Autonomous Driving. In Proceedings of the Conference on Robot Learn- ing, volume 155 of Proceedings of Machine Learning Re- search, 232–246. Yang, H.; Zhang, S.; Huang, D.; Wu, X.; Zhu, H.; He, T.; Tang, S.; Zhao, H.; Qiu, Q.; Lin, B.; He, X.; and Ouyang, W. 2024a. UniPAD: A Universal Pre-training Paradigm for Autonomous Driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 15238–15250. Yang, Z.; Chen, L.; Sun, Y.; and Li, H. 2024b. Visual Point Cloud Forecasting Enables Scalable Autonomous Driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 14673–14684. Zhang, H.; Zhou, W.; Zhu, Y.; Yan, X.; Gao, J.; Bai, D.; Cai, Y.; Liu, B.; Cui, S.; and Li, Z. 2025. VisionPAD: A Vision- Centric Pre-training Paradigm for Autonomous Driving. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, 17165–17175. Zhang, Q.; Peng, Z.; and Zhou, B. 2022. Learning to Drive by Watching YouTube Videos: Action-Conditioned Contrastive Policy Pretraining. In European Conference on Computer Vision, 111–128. Springer. Zheng, W.; Song, R.; Guo, X.; Zhang, C.; and Chen, L. 2024. GenAD: Generative End-to-End Autonomous Driv- ing. In European Conference on Computer Vision, 87–104. Springer. Zhou, S.; Liu, W.; Hu, C.; Zhou, S.; and Ma, C. 2023. Uni- Distill: A Universal Cross-Modality Knowledge Distillation Framework for 3D Object Detection in Bird’s-Eye View. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, 5116–5125. Zhu, Y.; Xue, Y.; Zhang, H.; Jiang, G.; Zhou, W.; Yan, X.; Gao, J.; Cai, Y.; Liu, B.; Li, Z.; and Shen, S. 2026. DLWM: Dual Latent World Models Enable Holistic Gaussian-Centric Pre-training in Autonomous Driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 39713–39723. Zou, J.; Liao, B.; Zhang, Q.; Liu, W.; and Wang, X. 2025. MIM4D: Masked Modeling with Multi-View Video for Au- tonomous Driving Representation Learning. International Journal of Computer Vision, 133: 6074–6087. Planning-Aligned Pretraining of BEV Representations with Sparse Action-Conditioned Targets for End-to-End Autonomous Driving Technical Supplement Method and Reproducibility Details This section records the implementation details needed to reproduce PAVER. Every value below is supported by the canonical configuration, implementation, or preserved exper- iment artifacts. Metric abbreviations follow their task-native evaluators. For perception, mAP is mean average precision and NDS the nuScenes detection score; for motion, EPA is the end-to-end prediction accuracy, ADE and FDE the aver- age and final displacement errors, and MR the miss rate; for occupancy, IoU and mIoU are the completion and semantic intersection over union. The analyses additionally use the area under the receiver operating characteristic (AUROC). The closed-loop results use the Bench2Drive driving score (DS), route completion (RC) and infraction score (IS). Model names follow one convention throughout this sup- plement. VAD-Tiny (scratch) is the reproduced VAD-Tiny trained end to end for 60 downstream epochs with no pre- training; it is the reference every VAD-Tiny comparison is made against, and tables abbreviate it to Scratch. VAD-Tiny (staged) is the same architecture given 48 epochs of pre- training on the downstream tasks followed by 12 fine-tuning epochs, so its 60-epoch total matches scratch and isolates the effect of the pretraining objective rather than of extra opti- mization; tables abbreviate it to Staged. + PAVER denotes 20 pretraining epochs with the PAVER objective followed by downstream fine-tuning, with only the BEV encoder trans- ferred. Where the auxiliary capacity varies, the parenthesized figure is the number of trainable parameters in the tempo- rary PAVER head. PAVER (10K) is the canonical point- wise readout; PAVER (30K) widens its projector and action- conditioned MLP, while PAVER (90K) replaces the wider pointwise projector with a 3× 3 projector that accesses local BEV context. The exact counts are given in Table 11. VAD- Base (scratch), GenAD (scratch) and UniAD-Tiny (scratch) are used in the same sense for the other architectures. The experiments use the VAD BEV model (Jiang et al. 2023) on nuScenes (Caesar et al. 2020); its spatially indexed BEV follows the multi-camera representation used by BEV- Former (Li et al. 2022). Model and Data Contract Table 10 records the data and model configuration. All camera pipelines use BGR normalization with mean (103.530, 116.280, 123.675), unit standard deviation, and no RGB conversion. The single-frame pretraining queue pre- vents the objective from using a recurrent BEV state. Detec- tion, mapping, motion-prediction, and planning decoders are disabled during pretraining. Rule-Based Action Sampling Let a s and ω s denote the perturbed acceleration and yaw rate for action slot s. Starting from current speed v 0 , the rollout is v s,t = max(0,v 0 + a s t∆t),(6) ψ s,t = ω s t∆t,(7) x s,t = t X j=1 v s,j cos(ψ s,j )∆t,(8) y s,t = t X j=1 v s,j sin(ψ s,j )∆t,(9) where ∆t = 0.5 s. Target construction and BEV indexing use the calibrated sensor frame, while the prediction head receives the corresponding ego-frame state [x,y,ψ,v]. Sparse Evidence and Action Targets Following occupancy-grid mapping (Elfes 1989), the evi- dence map is initialized as Unknown. For each valid LiDAR return, samples along the open ray segment are written as Free and the measured endpoint is then written as Occu- pied. At each action state, five samples across the vehicle width query this raster. The resulting soft ratios are super- vision targets, not a dense semantic-occupancy prediction. An observed-free cross section has target (r,u) = (0, 0) and remains part of the loss. Formally, for heading ̄ ψ s,t the lateral unit vector is ℓ( ̄ ψ s,t ) = (− sin ̄ ψ s,t , cos ̄ ψ s,t ), so the k-th query position is p s,t,k = ( ̄x s,t , ̄y s,t ) + d k ℓ( ̄ ψ s,t ),(10) with d k denoting the k-th symmetric lateral offset. Writing V s,t for the queries that survive the range check, the risk and unknown targets used in the main paper are the indicator averages r s,t = 1 |V s,t | X k∈V s,t ⊮ [E(p s,t,k ) = Occupied], (11) u s,t = 1 |V s,t | X k∈V s,t ⊮ [E(p s,t,k ) = Unknown]. (12) Out-of-range queries are dropped rather than clamped, so an action that leaves the mapped region contributes fewer samples instead of a biased target. This construction compresses the privileged sweep into two quantities for each action–horizon query: whether mea- sured endpoints intersect the vehicle width and whether ray evidence supports that region. Unlike objectives that recon- struct appearance, depth, or dense scene content (Yang et al. 2024a; Zou et al. 2025; Zhang et al. 2025), it does not ask the auxiliary branch to reproduce details that the downstream planner never queries directly. The prediction problem is small, but its coordinates are not: answering it requires the camera BEV to place free, occupied, and unsupported evi- dence at the locations indexed by an ego motion. ItemConfiguration Training split28,130 nuScenes keyframes Validation split6,019 nuScenes keyframes Camera input6 synchronized surround views Image backbone and neckResNet-50 and FPN BEV representation[B, 256, 100, 100] BEV rangex∈ [−15, 15) m, y ∈ [−30, 30) m Pretraining queue1 frame Privileged inputCurrent LiDAR sweep, used only for targets Active pretraining outputsRisk and Unknown logits Task annotationsNone Downstream inferenceCamera-only Table 10: PAVER data and model configuration. Learned component10K30K90K Mask token256256256 Projector4,112 8,224 73,760 First readout layer5,376 18,944 18,944 Output layer514 1,026 1,026 Total10,258 28,450 93,986 Table 11: Auxiliary parameter counts for the PAVER capacity variants. The 10K and 30K projectors are pointwise; the 90K projector uses a 3× 3 kernel. PAVER Head and Objective The mask M ∈ 0, 1 B×1×H×W is the union of all raster- ized lateral-query cells, and a cell reached by several queries is included once. With the shared token m∈ R 1×C×1×1 , the temporary copy is F mask = (1− M )⊙ F + M ⊙ m, the pointwise projector gives Z = φ(F mask ) ∈ R B×C p ×H×W , and each action–horizon entry reads a single feature z s,t = S(Z, ( ̄x s,t , ̄y s,t ))∈ R C p through bilinear samplingS (Jader- berg et al. 2015). The predictor input is the concatenation h s,t = [z s,t ;q ego s,t ], so the PAVER head sees a projected BEV feature whose corridor cells have been replaced, together with the action state that selected it. The action corridor is replaced only in the temporary BEV copy consumed by the PAVER head. The original BEV rep- resentation remains unchanged. The objective is L PAVER = BCELogit(ˆr,r) + 0.5 BCELogit(ˆu,u). (13) The 10,258-parameter PAVER head is therefore only a pre- diction head, not the representation being transferred. Mask- ing removes the feature at the supervised corridor cell, and each loss is differentiated through bilinear BEV sampling and the shared BEV encoder. Increasing predictor capacity is unnecessary for expressing the two targets; useful scene in- formation must instead be organized in the BEV encoder that remains after pretraining. This differs from masked recon- struction, where decoder capacity participates in recovering a high-dimensional observation (Devlin et al. 2019; Lin et al. 2024; Min et al. 2022). SettingValue Pretraining duration20 epochs Pretraining batch4 GPUs, 4 samples per GPU OptimizerAdamW (Loshchilov and Hutter 2019) Learning rate5× 10 −5 Image-backbone multiplier 0.2 Weight decay0.01 ScheduleCosine annealing (Loshchilov and Hutter 2017) Warmup500 iterations, linear, ratio 1/3 Minimum learning-rate ra- tio 10 −3 Gradient clippingNorm 35 Checkpoint interval1 epoch Downstream learning rate2× 10 −4 Downstream queue3 frames Table 12: Optimization settings for the reported PAVER ex- periments. Optimization and Transfer Under the settings of Table 12, all transferred image- backbone, FPN, and BEV-encoder parameters are optimized during pretraining. Afterward, the action sampler, target builder, mask token, projector, and prediction MLP are dis- carded. The transferred initialization covers the image back- bone, FPN, BEV queries and positional encodings, cam- era and feature-level embeddings, CAN-bus MLP, and BEV transformer encoder. Task decoders are newly instantiated, and downstream training starts from epoch 1 with a fresh optimizer and scheduler. No PAVER-specific parameter is active at inference. Frozen-Predictor Controls We evaluate the frozen auxiliary predictor on 192 nuScenes validation samples: Table 13 reports the interventions, Ta- ble 14 their scene-bootstrap uncertainty, Table 15 the off- support probe, and Figure 9 the action–time shortcut audit. Each intervention preserves the weights and targets while changing only the specified BEV or action-state input. These InputRisk BCE∆ Unknown BCE∆ Correct input0.3366 0.00000.2333 0.0000 Cross-scene BEV0.3571 0.02040.2723 0.0390 Zero context0.3511 0.01450.2784 0.0451 Spatial permutation 0.3691 0.03250.3254 0.0920 Action permutation 0.3879 0.05130.2688 0.0355 Table 13: Frozen-input intervention results. Lower is better; ∆ is measured from the correct-input result and computed before rounding the displayed losses. RiskUnknown 0.0 0.2 BCE (a) Prediction error RiskUnknown 0.0 0.2 (b) Residual signal Action-time prior Camera BEV Pearson r R 2 Figure 9: Action–time shortcut audit. The frozen BEV pre- dictor is compared with a prior conditioned on control family, horizon, and perturbed action state. controls diagnose input dependence; they are not downstream model-selection results. Every intervention increases both losses, and every paired interval excludes zero. Spatial permutation has the largest effect on Unknown, while action permutation has the largest effect on Risk. The predictor therefore uses the scene, its spatial arrangement, and the action–scene pairing rather than only a fixed action and horizon prior. The BEV predictor reduces Risk BCE by 0.0042 and Un- known BCE by 0.0061 relative to the action–time prior. Residual R 2 is below 0.05 for both targets, while Pearson correlations of 0.2410 and 0.3321 retain measurable scene dependence. Thus, action and horizon explain substantial tar- get variation, but do not fully explain the learned prediction. The sparse action targets occupy 2.66% of the BEV and re- main in the forward half of the ego frame. Identical frozen lin- ear probes nevertheless show AUROC changes of +0.0868 in rear regions, +0.0617 in lateral regions, and +0.1488 in front off-path regions; all paired 95% confidence intervals exclude zero. These measurements establish off-support rep- resentation change, not direct supervision at those cells. The architecture exposes a specific path for this change. All six views pass through the same ResNet and FPN. Camera identity and calibration then determine where each view is sampled, but the deformable-attention transformations and output projection are shared before visible-view contribu- tions are accumulated in a common BEV (Li et al. 2022; Jiang et al. 2023). A sparse loss therefore updates parame- ters that are reused across cameras and BEV locations even Action corridorAdjacentOff corridor 0.55 0.60 0.65 0.70 0.75 0.80 0.85 0.90 Frozen evidence probes Initialization PAVER (pretrained) PAVER Baseline (scratch) Figure 10: Frozen evidence probes by BEV region. PAVER pretraining increases linear accessibility across regions; downstream differences are smaller and mixed. though only a small set of output cells receives an immediate gradient. This distinction separates spatially sparse targets from a camera-specific or head-only optimization path. This mechanism explains how an action-aligned objective can influence the BEV later consumed by detection and map- ping without pretraining their decoders. Both perception end- points improve on VAD-Tiny: map mAP rises from 0.4188 to 0.4386 and detection mAP from 0.2306 to 0.2784. Transfer is not uniform across architectures, so the evidence supports broad representational reach rather than direct learning of task semantics. Comprehensive Analysis The endpoint comparisons establish transfer performance but do not explain what the pretraining objective learns or how its effect develops during downstream optimization. Unless stated otherwise, downstream analyses use the selected VAD- Tiny PAVER model reported in the main experiments. The analyses test a specific account of PAVER’s efficiency: dense objectives preserve broad appearance and 3D scene content (Yang et al. 2024a; Zou et al. 2025; Zhang et al. 2025), whereas planning consumes spatial evidence through candidate ego motions (Hu et al. 2023; Jiang et al. 2023), and PAVER retains only their intersection. The auxiliary head can then remain small, because the shared BEV encoder is what must place the relevant evidence at the queried locations. Changes in the Transferred BEV Representation Matched frozen linear probes predict Unknown, Free, and Occupied evidence from camera-derived BEV features (Fig- ure 10). We use ROC-AUC to measure linear accessibility Cross-sceneZero ctx.Spatial perm.Action perm. ∆ Risk [95% CI]0.0204 [0.0150, 0.0241]0.0145 [0.0099, 0.0176]0.0325 [0.0292, 0.0358]0.0513 [0.0500, 0.0535] ∆ Unknown [95% CI]0.0390 [0.0288, 0.0464]0.0451 [0.0390, 0.0512]0.0920 [0.0835, 0.1002]0.0355 [0.0329, 0.0378] Table 14: Scene-bootstrap uncertainty of the frozen-input interventions. RegionInit. PAVER∆95% CI Rear0.6532 0.7399 +0.0868 [0.0598, 0.1131] Lateral0.6455 0.7072 +0.0617 [0.0250, 0.0956] Front off-path 0.6716 0.8205 +0.1488 [0.1032, 0.1916] Table 15: Frozen object–background probe AUROC out- side direct action support. Differences are computed before rounding the displayed AUROC values. independently of a single decision threshold (Fawcett 2006). PAVER pretraining increases macro ROC-AUC from 0.8484 to 0.8931 in the action corridor, from 0.7326 to 0.7755 in the adjacent ring, and from 0.6708 to 0.7175 off corridor. After downstream training, the all-region probe remains slightly above the reproduced scratch model (0.7692 versus 0.7618), but the regional result is mixed: corridor AUC is 0.8600 versus 0.8635, whereas adjacent and off-corridor AUC are 0.7573 versus 0.7532 and 0.7348 versus 0.7254. The tar- get is spatially sparse, but its representation changes are not confined to the supervised cells. This pattern clarifies the role of the pretraining-only auxil- iary head. Its 10,258 parameters are discarded, so they cannot supply downstream capacity. What persists is easier linear ac- cess to evidence where candidate actions query the BEV. The objective therefore does not need to encode a complete scene in its auxiliary output: it only needs enough capacity to ex- pose whether the shared BEV has organized local geometric support around an action. Comparable probe gains outside the corridor are consistent with updating the shared image and BEV modules used to construct the full representation, rather than adding capacity only at queried cells. These frozen probes measure linear accessibility, not causal use by downstream decoders. Downstream Adaptation and Checkpoint Dynamics Figure 11 shows 13 VAD-Tiny experiment series. Each line uses native per-checkpoint evaluator outputs only; missing epochs are not interpolated. PAVER (10K) is emphasized for readability. Detection and map metrics tend to improve broadly, whereas planning and collision are noisier and non- monotonic. The trajectories are consistent with representa- tion refinement together with checkpoint and safety sensitiv- ity, not with causality or universal superiority. All 252 selected parameter tensors change during pre- training (Figure 12). Relative L2 change is 0.1017 for the BEV Encoder, 0.0845 for the Image Backbone, and 0.2215 for the Image Neck. The Image Backbone retains high di- rectional similarity (0.9964 cosine similarity) but has no unchanged parameter elements. This agrees with the train- ing configuration, which leaves the backbone trainable with a 0.2 learning-rate multiplier. The measurements establish weight-space change, not that any one component causes the downstream gains. Rapid early drift is compatible with useful pretraining rather than evidence that the initialization is immediately for- gotten. The encoder is free to move because the pretraining- only auxiliary head imposes no downstream architectural constraint; what matters is that optimization begins from a BEV in which action-relevant evidence is already accessible. Figure 13 shows that, at downstream epoch 30, relative L2 changes from the pretrained state are 0.3737 for the BEV Encoder, 0.1913 for the Image Backbone, and 1.0992 for the Image Neck. From epochs 20 to 30, these values increase by only 0.0176, 0.0106, and 0.0485, respectively. Thus, the transferred parameters remain trainable, while most of their weight-space movement occurs earlier in downstream opti- mization. Metric Structure The first three within-series metric components explain 92.30% of trajectory variance, but collision loads on a sep- arate component and correlates only weakly with the other utilities. Collision must therefore remain an explicit safety axis rather than being absorbed into one balanced score. The separate collision component helps explain why trans- fer is not uniform across downstream architectures. PAVER supervises current geometric evidence, not a future collision event; converting that evidence into a trajectory still depends on the planner and its interaction with motion and occupancy modules (Hu et al. 2022, 2023; Jiang et al. 2023). Better BEV organization can therefore improve displacement and per- ception without guaranteeing the same collision response for every decoder. We therefore retain collision explicitly so that this decoder dependence is not hidden inside an aggregate score. Planning Error Tails Averages hide where a planner actually fails, so we rank val- idation samples by their own average L2 and report the mean over the worst tenth, twentieth and hundredth (Table 16). The margin PAVER holds over each baseline widens with sever- ity on all three architectures: on VAD-Tiny it grows from 0.06 m across all samples to 3.47 m within the worst 1%, on VAD-Base from 0.18 to 3.55 m, and on GenAD from 0.05 to 1.86 m. The overall column reproduces the planning column of the multi-task table, so the tails come from the same checkpoints rather than from a separate evaluation. Auxiliary Target Quality at the End of Pretraining We evaluate each auxiliary head at pretraining epoch 20 on all 6,019 validation frames. Table 17 keeps task-native met- 0.6 0.8 1.0 1.2 1.4 1.6 1.8 Planning L2 Avg 0.75 1.00 1.25 1.50 1.75 2.00 2.25 2.50 Planning L2 3s 0.2 0.4 0.6 0.8 1.0 1.2 1.4 Box Collision Avg (%) 0.8 1.0 1.2 1.4 1.6 1.8 Motion ADE 1.00 1.25 1.50 1.75 2.00 2.25 2.50 Motion FDE 0.10 0.12 0.14 0.16 0.18 0.20 0.22 0.24 0.26 Motion MR 102030405060 Downstream epoch 0.10 0.15 0.20 0.25 Detection mAP 102030405060 Downstream epoch 0.10 0.15 0.20 0.25 0.30 0.35 0.40 Detection NDS 102030405060 Downstream epoch 0.10 0.15 0.20 0.25 0.30 0.35 0.40 0.45 Map mAP PAVER (10K) PAVER (30K) PAVER (90K) PAVER + Pseudo-LiDAR Scratch Staged w/o Mask w/o Action State Detection Map Occupancy Detection + Map Detection + Map + Occupancy PAVER (10K) PAVER (30K) PAVER (90K) PAVER + Pseudo-LiDAR Scratch Staged w/o Mask w/o Action State Detection Map Occupancy Detection + Map Detection + Map + Occupancy Figure 11: Native per-checkpoint evaluator trajectories for 13 VAD-Tiny experiments. Missing epochs are not interpolated; PAVER (10K) is emphasized. 05101520 Pretraining epoch (0 = before first batch) 0 2 4 6 8 L2 norm Mask-token trajectory Mask token Projected mask 0.000.050.100.150.20 Relative L2 change BEV Encoder BEV Position BEV Query Camera Embeddings CAN-bus MLP Image Backbone Image Neck Level Embeddings Change during pretraining 0.00.20.40.60.81.0 Relative L2 change BEV Encoder BEV Position BEV Query Camera Embeddings CAN-bus MLP Image Backbone Image Neck Level Embeddings Change during downstream training Figure 12: Parameter dynamics during PAVER pretraining and downstream training. The panels show the learned mask-token trajectory, changes from the state before the first pretraining batch to pretraining epoch 20, and subsequent changes at downstream epoch 30. Registered buffers are excluded. Relative L2 change 0 0.1 0.2 0.3 0.4 0102030 Fine-tuning epoch Cosine similarity 0.92 0.94 0.96 0.98 1 0102030 Fine-tuning epoch BEV EncoderBEV PositionBEV Query Camera EmbeddingsCAN-bus MLPImage Backbone Figure 13: Parameter-only changes from PAVER pretraining to downstream epoch 30 for six selected parameter groups. Left: L2 distance divided by the pretrained parameter norm. Right: cosine similarity to the pretrained parameters. Both panels share the centered legend; epoch 0 is the pretrained reference. Image-neck changes are reported in the text. L2 (m)↓ All Worst 10% Worst 5% Worst 1% VAD-Tiny † 0.662.223.096.98 + Ours0.601.712.153.51 VAD-Base 0.742.854.2810.56 + Ours0.562.083.057.01 GenAD † 0.592.102.986.16 + Ours0.541.732.304.30 † Reproduced results. Table 16: Planning results by error percentile on nuScenes. Samples are ranked by their own average L2; each column reports the mean over that fraction of the 5,119 valid planning samples. Pretraining target Det. mAP NDS Map mAP Occ. IoU Occ. mIoU Detection0.1208 0.1907 – Map– 0.0040 – Occupancy– 0.6011 0.2045 Det. + Map + Occ. 0.1314 0.1771 0.0120 0.5579 0.1907 Table 17: Auxiliary-head quality after 20 pretraining epochs. Det., Map, and Occ. denote pretraining on detection, on the vector map, and on occupancy, so Det. + Map + Occ. is the joint objective. Each column uses its task-native evaluator. rics separate: nuScenes detection mAP and NDS, vector-map mAP, and occupancy completion IoU and semantic mIoU. A AbsRel↓ RMSE↓ SILog↓ δ 1 ↑ δ 2 ↑ δ 3 ↑ 0.1286 5.5592 21.7424 0.9238 0.9593 0.9739 Table 18: Pseudo-LiDAR depth quality on a 1,000-frame training audit. RMSE is in meters. The 95% intervals are [0.1268, 0.1305] for AbsRel and [0.9225, 0.9250] for δ 1 . dash means that the corresponding head is absent, not that its score is zero. These numbers report whether each temporary head learned its own target at the end of pretraining. The combined objective improves detection mAP from 0.1208 to 0.1314 and map mAP from 0.0040 to 0.0120 rela- tive to the corresponding single-task heads, but its detection NDS decreases from 0.1907 to 0.1771. Occupancy com- pletion IoU and semantic mIoU also decrease from 0.6011 and 0.2045 to 0.5579 and 0.1907. The mixed signs rule out the simpler explanation that adding more auxiliary tasks uni- formly improves pretraining quality. They are consistent with competition between task losses, but do not isolate gradient interference as the cause. Target Quality and Qualitative Evidence Pseudo-LiDAR Target Quality The pseudo-LiDAR audit measures the metric-depth source used by the UniK3D variant before any downstream training. We draw 1,000 training frames by scene-stratified sampling, evaluate all six cameras, and pool projected points within each frame. Predictions are evaluated in camera-depth co- ordinates over 1–80 m and within a 20 m LiDAR radius. No scale or shift alignment is applied. Table 18 therefore ModelMethod / PretrainingDivider Ped. Crossing Boundary mAP VAD-Tiny Scratch0.4760.3170.463 0.419 VAD-Tiny Staged0.4450.3520.443 0.413 VAD-Tiny PAVER (10K)0.4830.3390.494 0.439 VAD-Tiny PAVER (30K)0.4800.3830.479 0.447 VAD-Tiny PAVER (90K)0.4690.3290.501 0.433 VAD-Tiny w/o Mask0.4790.3720.484 0.445 VAD-Tiny w/o Action State0.4650.3750.488 0.443 VAD-Tiny Pseudo-LiDAR0.4750.3780.465 0.439 VAD-Tiny Detection0.4440.3360.437 0.406 VAD-Tiny Map0.4020.3360.434 0.391 VAD-Tiny Occupancy0.4290.3180.434 0.393 VAD-Tiny Detection + Map0.4230.3290.438 0.397 VAD-Tiny Detection + Map + Occupancy 0.4390.3720.450 0.420 VAD-Base Scratch0.5300.4400.535 0.502 VAD-Base PAVER0.5320.4380.523 0.498 GenADScratch0.5010.3760.488 0.455 GenADPAVER0.4730.3540.484 0.437 Table 20: Vector map prediction results, aggregate and class-wise. evaluates the raw metric predictions rather than a fitted vi- sualization. The audit contains 6,000 camera images and 15,425,270 valid projected points. The highδ 1 coverage shows that most projected points pre- serve useful metric scale, while the larger RMSE and SILog expose non-negligible long-tail depth errors. This distinction matters for interpretation: pseudo-LiDAR can provide broad geometric support without being equivalent to measured Li- DAR. Figures 21 and 22 compare RGB-colored measured and pseudo point clouds from the same camera frames in top-down and oblique views. We rank all audited frames by δ 1 , retain the top 30, and choose 10 from that fixed pool to cover distinct road layouts, lighting conditions, and traffic compositions. Capacity and Pretraining-Target Variants Table 19 reports the fixed checkpoints used for the capacity and target comparisons, Figure 14 plots auxiliary capacity against every native metric, and Figures 36, 37, 38 and 39 show the corresponding downstream predictions. Increasing the auxiliary capacity does not produce a monotonic multi- task gain. PAVER (30K) gives the lowest planning L2 and the highest map mAP in this group. PAVER (90K) has the best ADE and detection mAP, whereas PAVER (10K) has the lowest collision and the highest exact NDS (0.3968 versus 0.3966 for PAVER (90K), although both display as 0.397). Pseudo-LiDAR and joint detection–map–occupancy targets also reduce L2 relative to scratch training, but their motion, collision, and perception metrics move in different directions. These mixed signs rule out a simple explanation based on auxiliary capacity or target complexity alone. Cross-Model Transfer Evidence Table 19 reports the fixed VAD-Tiny, VAD-Base, and GenAD checkpoints used in the main paper, and Figures 40 and 41 show the VAD-Base and GenAD predictions with and with- out PAVER initialization. On VAD-Tiny, PAVER improves planning, collision, motion, detection, NDS, and map mAP. On VAD-Base, PAVER reduces average L2, improves motion and detection, raises collision, and leaves map mAP essen- tially unchanged at the displayed precision. On GenAD, it improves planning, collision, motion, detection, and NDS, but map mAP decreases. The different metric signs are con- sistent with a shared BEV initialization whose downstream effect remains architecture dependent; they do not support a uniform improvement claim. Class-wise Perception Results Table 21 reports every available nuScenes detection class AP for the fixed checkpoints used in the component, capac- ity, auxiliary-target, pseudo-LiDAR, and cross-model com- parisons. The VAD-Tiny (scratch) values are taken from the preserved native epoch-60 detection summary, including its per-class evaluator output. For every complete row, the mean of the displayed class AP values agrees with the reported mAP within 0.0006; the residual difference is source round- ing. The class-level results show redistribution rather than uni- form gains. PAVER (90K) gives the strongest VAD-Tiny aggregate detection result, while component removal and auxiliary-target changes favor different object classes. Map- only pretraining is weakest in aggregate detection. VAD-Base changes are mixed, whereas GenAD improves aggregate de- tection and every class with nonzero AP under PAVER ini- tialization. Table 20 applies the same fixed-checkpoint policy to vector-map prediction. The VAD-Tiny (scratch) class val- ues come from the preserved native epoch-60 map evalu- ation. Increasing the temporary PAVER capacity does not yield a monotonic map trend: PAVER (30K) has the highest PlanningMotionDetectionMap ModelMethod / PretrainingL2↓ Col. (%)↓ EPA↑ ADE↓ FDE↓ MR↓ mAP↑ NDS↑ mAP↑ VAD-Tiny Scratch0.6620.5130.458 0.905 1.250 0.135 0.231 0.338 0.419 VAD-Tiny Staged0.7010.7300.455 0.840 1.151 0.122 0.232 0.363 0.413 VAD-Tiny PAVER (10K)0.6030.1870.498 0.803 1.086 0.121 0.278 0.397 0.439 VAD-Tiny PAVER (30K)0.5140.3200.492 0.800 1.090 0.125 0.256 0.383 0.447 VAD-Tiny PAVER (90K)0.5960.2200.509 0.782 1.056 0.120 0.280 0.397 0.433 VAD-Tiny w/o Mask0.7050.5500.499 0.797 1.080 0.123 0.276 0.388 0.445 VAD-Tiny w/o Action State0.6550.5230.508 0.808 1.084 0.115 0.274 0.396 0.443 VAD-Tiny Pseudo-LiDAR0.5630.5600.495 0.831 1.101 0.121 0.268 0.386 0.439 VAD-Tiny Detection0.5150.2400.455 0.798 1.067 0.119 0.254 0.365 0.406 VAD-Tiny Map0.6940.3330.429 0.857 1.153 0.123 0.213 0.331 0.391 VAD-Tiny Occupancy0.6460.3200.438 0.905 1.230 0.136 0.219 0.336 0.393 VAD-Tiny Detection + Map0.5040.1900.432 0.805 1.073 0.117 0.234 0.361 0.397 VAD-Tiny Detection + Map + Occupancy 0.5770.2700.457 0.798 1.082 0.122 0.243 0.356 0.420 VAD-Base Scratch0.7420.3070.562 0.760 1.027 0.111 0.291 0.422 0.502 VAD-Base PAVER0.5620.4010.592 0.685 0.902 0.093 0.327 0.446 0.498 GenADScratch0.5910.3700.413 0.868 1.184 0.140 0.188 0.263 0.455 GenADPAVER0.5370.2070.462 0.800 1.053 0.114 0.210 0.277 0.437 Table 19: Planning, motion, detection, and map results on nuScenes. Planning values are 3-horizon averages, collision is in percent, and motion uses the 2.0 m matching threshold. ModelMethod / PretrainingmAP NDS Car Truck C.Veh. Bus Trail. Barr. Motor. Bicy. Ped. Cone VAD-Tiny Scratch0.231 0.338 0.492 0.189 0.060 0.283 0.005 0.338 0.126 0.171 0.275 0.367 VAD-Tiny Staged0.232 0.363 0.485 0.168 0.059 0.338 0.004 0.336 0.126 0.162 0.299 0.345 VAD-Tiny PAVER (10K)0.278 0.397 0.509 0.237 0.109 0.414 0.010 0.323 0.217 0.227 0.346 0.391 VAD-Tiny PAVER (30K)0.256 0.383 0.514 0.245 0.060 0.335 0.009 0.298 0.213 0.215 0.334 0.335 VAD-Tiny PAVER (90K)0.280 0.397 0.527 0.220 0.099 0.369 0.013 0.318 0.258 0.237 0.335 0.419 VAD-Tiny w/o Mask0.276 0.388 0.506 0.217 0.081 0.394 0.012 0.316 0.236 0.268 0.344 0.388 VAD-Tiny w/o Action State0.274 0.396 0.513 0.213 0.130 0.332 0.009 0.312 0.248 0.243 0.340 0.399 VAD-Tiny Pseudo-LiDAR0.268 0.386 0.503 0.227 0.081 0.323 0.018 0.303 0.232 0.281 0.340 0.377 VAD-Tiny Detection0.254 0.365 0.500 0.197 0.041 0.359 0.005 0.334 0.157 0.243 0.315 0.385 VAD-Tiny Map0.213 0.331 0.475 0.180 0.043 0.228 0.005 0.289 0.155 0.182 0.254 0.322 VAD-Tiny Occupancy0.219 0.336 0.457 0.189 0.033 0.240 0.007 0.289 0.122 0.214 0.297 0.339 VAD-Tiny Detection + Map0.234 0.361 0.502 0.146 0.061 0.314 0.003 0.312 0.171 0.176 0.284 0.369 VAD-Tiny Detection + Map + Occ. 0.243 0.356 0.513 0.184 0.078 0.260 0.004 0.298 0.170 0.246 0.301 0.379 VAD-Base Scratch0.291 0.422 0.548 0.240 0.081 0.332 0.017 0.358 0.244 0.267 0.378 0.442 VAD-Base PAVER0.327 0.446 0.550 0.243 0.124 0.460 0.029 0.385 0.311 0.293 0.423 0.452 GenADScratch0.188 0.263 0.443 0.000 0.077 0.000 0.000 0.298 0.169 0.224 0.290 0.373 GenADPAVER0.210 0.277 0.459 0.000 0.102 0.000 0.000 0.314 0.226 0.249 0.342 0.412 Table 21: 3D object detection results, aggregate and class-wise. C.Veh., Trail., Barr., Motor., Bicy., Ped., and Cone denote construction vehicle, trailer, barrier, motorcycle, bicycle, pedestrian, and traffic cone. A dash marks a class the evaluator did not report. VAD-Tiny map mAP among the three capacities, whereas PAVER (90K) is lower than PAVER (10K). The component and auxiliary-target rows likewise move the three map classes differently, which argues against attributing transfer to uni- formly better vector-map reconstruction. Figures 23–27, 28, and 29 show what each auxiliary head reconstructs at the end of its 20-epoch budget. Component Ablation Metrics The component rows of Table 19 carry the aggregate met- rics for this comparison, and Figures 34 and 35 show the two ablations against the selected VAD-Tiny model of Fig- ures 32, 33, 30 and 31. Removing action conditioning keeps the action-corridor mask, whereas removing the mask retains action conditioning. Full PAVER has the lowest planning L2 and collision at every horizon, as well as the highest exact 10 4 10 5 10 6 10 7 0.5 0.6 0.7 0.8 Planning L2 (m) ↓ VAD-Tiny Planning L2 (m) PAVER (10K) PAVER (30K) PAVER (90K) Det. Map Occupancy Det. + Map Det. + Map + Occ. UniPAD MIM4D-T 10 4 10 5 10 6 10 7 0.2 0.3 0.4 0.5 Collision (%) ↓ VAD-Tiny Collision (%) PAVER (10K) PAVER (30K) PAVER (90K) Det. Map Occupancy Det. + Map Det. + Map + Occ. UniPAD MIM4D-T 10 4 10 5 10 6 10 7 0.32 0.34 0.36 0.38 0.40 Detection NDS ↑ VAD-Tiny Detection NDS PAVER (10K) PAVER (30K) PAVER (90K) Det. Map Occupancy Det. + Map Det. + Map + Occ. 10 4 10 5 10 6 10 7 0.20 0.22 0.24 0.26 0.28 Detection mAP ↑ VAD-Tiny Detection mAP PAVER (10K) PAVER (30K) PAVER (90K) Det. Map Occupancy Det. + Map Det. + Map + Occ. 10 4 10 5 10 6 10 7 0.38 0.40 0.42 0.44 0.46 Map mAP ↑ VAD-Tiny Map mAP PAVER (10K) PAVER (30K) PAVER (90K) Det. Map Occupancy Det. + Map Det. + Map + Occ. 10 4 10 5 10 6 10 7 0.42 0.44 0.46 0.48 0.50 0.52 Motion EPA ↑ VAD-Tiny Motion EPA PAVER (10K) PAVER (30K) PAVER (90K) Det. Map Occupancy Det. + Map Det. + Map + Occ. 10 4 10 5 10 6 10 7 Additional pretraining parameters (log scale) 0.775 0.800 0.825 0.850 0.875 0.900 0.925 Motion ADE (m) ↓ VAD-Tiny Motion ADE (m) PAVER (10K) PAVER (30K) PAVER (90K) Det. Map Occupancy Det. + Map Det. + Map + Occ. 10 4 10 5 10 6 10 7 Additional pretraining parameters (log scale) 1.05 1.10 1.15 1.20 1.25 Motion FDE (m) ↓ VAD-Tiny Motion FDE (m) PAVER (10K) PAVER (30K) PAVER (90K) Det. Map Occupancy Det. + Map Det. + Map + Occ. 10 4 10 5 10 6 10 7 Additional pretraining parameters (log scale) 0.115 0.120 0.125 0.130 0.135 0.140 Motion MR ↓ VAD-Tiny Motion MR PAVER (10K) PAVER (30K) PAVER (90K) Det. Map Occupancy Det. + Map Det. + Map + Occ. PAVER (10K) PAVER (30K) PAVER (90K) Task supervision Reported baseline VAD-Tiny scratch Figure 14: Auxiliary capacity against every native metric. PAVER, the task-supervised controls, and the reported UniPAD and MIM4D baselines are shown against the VAD-Tiny scratch reference. NDS (0.3968). The partial variants retain better individual motion or map metrics: removing action conditioning im- proves FDE, MR, EPA, and map mAP, while removing the mask improves ADE, FDE, EPA, and map mAP. Because each removal costs planning accuracy and safety while trad- ing away a different auxiliary metric, both components are load-bearing. Closed-Loop Route Outcomes on Bench2Drive Table 22 audits every Bench2Drive Town05 Long route be- hind the closed-loop results of the main paper, with one repe- tition per route. Timeout marks an exhausted 200 s simulated- time budget. PAVER initialization converts three UniAD- Tiny timeouts into completions. Figure 15 shows every route as a timeline. Three UniAD-Tiny routes carry a frame-by- frame figure: the two whose outcome changes and whose baseline frames were recorded, and one where both policies stop but at very different points. Routes 26956, 26966 and 27506 recorded no baseline frames, so they appear in the audit only. The GenAD policy ran the same nine routes and two of them are shown frame by frame as well. The three filmstrips show a shared baseline failure mode: the policy reaches the scenario obstacle, slows, and does not resume. With PAVER initialization, UniAD-Tiny passes the parked obstacle at night (route 25318) and the accident blocking the oncoming lane (route 25857), completing both routes. In the parking cut-in scenario (route 24759), both policies remain blocked, but PAVER increases route comple- tion from 25.42% to 66.91% (Table 22). The two completed routes require manoeuvres around static blockers rather than lane following, consistent with the larger open-loop gains in the error tail than at the median. Figures 16, 17 and 18 show these three cases frame by frame. Figures 19 and 20 show the GenAD policy on routes 24759 and 25955. Neither GenAD run has a recorded completion trace, so both are stepped in elapsed time rather than matched on route completion, and the campaign holds no scratch GenAD run to pair them with; they document how a second architecture behaves after the same pretraining rather than a controlled comparison. Route ScenarioPolicyOutcomeRC↑DS↑ Collisions↓ 24759 Parking cut inUniAD-TinyBlocked25.42 25.420 + PAVERBlocked66.9152.430 GenAD + PAVER Completed † 100.00 100.000 25318 Parked obstacleUniAD-TinyTimeout35.69 21.411 + PAVERCompleted100.0036.002 GenAD + PAVER Timeout † 31.86 31.860 25381 Hazard at side laneUniAD-TinyCompleted 100.00 100.000 + PAVERCompleted100.00100.000 GenAD + PAVER Completed † 100.00 51.091 25857 Accident two waysUniAD-TinyTimeout32.30 32.300 + PAVERCompleted100.00100.000 GenAD + PAVER Timeout † 33.83 33.830 25955 Hazard at side lane two ways UniAD-TinyCompleted 100.00 100.000 + PAVERCompleted100.0036.002 GenAD + PAVER Completed † 100.00 60.001 26396 Static cut inUniAD-TinyCompleted 100.00 100.000 + PAVERCompleted100.00100.000 GenAD + PAVER Completed † 100.00 100.000 26956 Signalized junction right turn UniAD-TinyTimeout26.72 26.720 + PAVERCompleted100.0060.001 GenAD + PAVER Timeout † 22.64 22.640 26966 Signalized junction right turn UniAD-TinyTimeout28.48 28.480 + PAVERTimeout18.1518.150 GenAD + PAVER Timeout † 20.73 20.730 27506 Enter actor flowUniAD-TinyCompleted 100.00 1.747 + PAVERTimeout26.5226.520 GenAD + PAVER Timeout † 21.50 21.500 Table 22: Bench2Drive Town05 Long, per-route outcome audit. One repetition per route; RC is route completion, DS the composed driving score. Timeout marks an exhausted 200 s simulated-time budget. † scored without a completion trace, so it cannot be matched on route completion. 0 20 40 60 80 100 Route completion (%) Route 24759 · Parking cut inRoute 25318 · Parked obstacleRoute 25381 · Hazard at side lane 0 20 40 60 80 100 Route completion (%) Route 25857 · Accident two waysRoute 25955 · Hazard at side lane two waysRoute 26396 · Static cut in 050100150200 Simulated time (s) 0 20 40 60 80 100 Route completion (%) Route 26956 · Signalized junction right turn 050100150200 Simulated time (s) Route 26966 · Signalized junction right turn 050100150200 Simulated time (s) Route 27506 · Enter actor flow UniAD-TinyUniAD-Tiny + PAVERGenAD + PAVERCollision (policy colour)CompletedBlockedTimeout (200 s) Figure 15: Bench2Drive Town05 Long per-route timelines. Curves show recorded route completion against simulated time; markers indicate collisions, completion, or timeout. A missing completion trace is not a zero-completion trajectory. Mechanistic Summary The analyses support a 4-part explanation for the observed transfer. First, PAVER compresses privileged LiDAR ev- idence into two ratios instead of reconstructing a high- dimensional scene. Second, rule-based actions determine where those ratios are queried, aligning the supervision with spatial evidence later used by planning. Third, super- vision is spatially sparse, but frozen probes detect repre- sentation changes beyond the sampled corridor. Fourth, the pretraining-only auxiliary head is removed, while the trans- ferred encoder remains adaptable; early downstream weight drift and multi-task learning curves show refinement rather than dependence on a retained pretraining decoder. This combination explains why parameter count and target difficulty need not track transfer quality. A large decoder is useful when pretraining must recover RGB, depth, dense oc- cupancy, or future scene state (Yang et al. 2024a; Zou et al. 2025; Min et al. 2024; Yang et al. 2024b). PAVER asks a smaller question whose coordinates carry the planning prior, so the 10K parameters only map an already shared BEV feature and action state to two outputs. The useful capacity is the downstream BEV encoder being organized, not the pretraining-only auxiliary head being enlarged. The same account also predicts the observed limit: because the targets summarize current geometric support, transfer can weaken when occluded intent or future agent dynamics dominate. Evidence Scope and Limitations The comprehensive analyses use preserved checkpoints and matched nuScenes samples. The native trajectories in Fig- ure 11 extend to the available checkpoints, including epoch 40 for pretraining variants and epoch 60 for scratch. Each configuration is trained once; scene-cluster bootstrap inter- vals measure variation across sampled scenes, not across independent training runs. Correlated checkpoints are not treated as independent trials. The frozen-predictor controls establish dependence on BEV content, spatial arrangement, and action state, but do not iso- late every architectural mechanism. The closed-loop routes are one repetition each on nine routes, which indicates trans- fer rather than a ranking. These results remain diagnostic evidence and should not be extrapolated to general driving safety. 4% of route t = 2.5 s ImageTop-down View t = 2.5 s ImageTop-down View 10% of route t = 4.0 st = 4.0 s 18% of route t = 6.0 st = 6.0 s 25% of route t = 8.0 st = 8.0 s 30% of route t = 9.5 st = 9.0 s Final state t = 199.5 s Timeout at 200 s · DS 32.30 t = 30.5 s Completed 100% · DS 100.00 UniAD-TinyUniAD-Tiny + PAVER 4% of route t = 2.5 s ImageTop-down View t = 2.5 s ImageTop-down View 10% of route t = 4.0 st = 4.0 s 18% of route t = 6.0 st = 6.0 s 25% of route t = 8.0 st = 8.0 s 30% of route t = 9.5 st = 9.0 s Final state t = 199.5 s Timeout at 200 s · DS 32.30 t = 30.5 s Completed 100% · DS 100.00 UniAD-TinyUniAD-Tiny + PAVER Figure 16: Bench2Drive Town05 Long, route 25857: accident two ways. Rows are matched on route completion, so both policies are shown at the same physical point; each camera image is paired with the simulator’s top-down view. UniAD-Tiny exhausts the 200 s budget at 32.30 route completion, while the same architecture initialized with PAVER completes the route (Table 22). 4% of route t = 2.5 s ImageTop-down View t = 2.5 s ImageTop-down View 10% of route t = 4.0 st = 4.0 s 18% of route t = 6.5 st = 6.5 s 25% of route t = 8.5 st = 8.5 s 30% of route t = 11.0 st = 11.0 s Final state t = 199.5 s Timeout at 200 s · DS 21.41 t = 68.5 s Completed 100% · DS 36.00 UniAD-TinyUniAD-Tiny + PAVER 4% of route t = 2.5 s ImageTop-down View t = 2.5 s ImageTop-down View 10% of route t = 4.0 st = 4.0 s 18% of route t = 6.5 st = 6.5 s 25% of route t = 8.5 st = 8.5 s 30% of route t = 11.0 st = 11.0 s Final state t = 199.5 s Timeout at 200 s · DS 21.41 t = 68.5 s Completed 100% · DS 36.00 UniAD-TinyUniAD-Tiny + PAVER Figure 17: Bench2Drive Town05 Long, route 25318: parked obstacle at night. UniAD-Tiny exhausts the time budget behind the obstacle; with PAVER initialization the same architecture passes it and reaches the end of the route. 4% of route t = 2.5 s ImageTop-down View t = 2.5 s ImageTop-down View 10% of route t = 5.0 st = 4.0 s 18% of route t = 7.0 st = 6.0 s 25% of route t = 10.5 st = 8.0 s 30% of route not reached t = 10.5 s Final state t = 70.5 s Blocked · DS 25.42 t = 81.5 s Blocked · DS 52.43 UniAD-TinyUniAD-Tiny + PAVER 4% of route t = 2.5 s ImageTop-down View t = 2.5 s ImageTop-down View 10% of route t = 5.0 st = 4.0 s 18% of route t = 7.0 st = 6.0 s 25% of route t = 10.5 st = 8.0 s 30% of route not reached t = 10.5 s Final state t = 70.5 s Blocked · DS 25.42 t = 81.5 s Blocked · DS 52.43 UniAD-TinyUniAD-Tiny + PAVER Figure 18: Bench2Drive Town05 Long, route 24759: parking cut in. Both policies are blocked by the cut-in vehicle, but the pretrained model travels more than twice as far before it stops (Table 22). Image 2.5 s Step 1 6.0 s Step 2 10.5 s Step 3 14.5 s Step 4 17.5 s Step 5 23.5 s Final state Top-down View Completed 100% DS 100.00 Figure 19: Bench2Drive Town05 Long, route 24759: parking cut in, for GenAD initialized with PAVER. The camera image is paired with the simulator’s top-down view, on which the planned path is drawn as a band running from its near end to its far end. This run has no completion trace, so the columns are spaced by elapsed time rather than by route completion, and the campaign contains no scratch GenAD counterpart, so the figure reports single-policy behaviour. The policy passes the cut-in vehicle and completes the route. Image 3.0 s Step 1 7.0 s Step 2 12.5 s Step 3 17.5 s Step 4 21.0 s Step 5 28.5 s Final state Top-down View Completed 100% DS 60.00 Figure 20: Bench2Drive Town05 Long, route 25955: hazard at side lane two ways, for GenAD initialized with PAVER. The policy passes a stopped vehicle and a cyclist in fog and reaches the end of the route; the driving score of 60.00 reflects one recorded collision on an otherwise completed route (Table 22). Top-down ViewOblique View Input Images LiDAR Pseudo-LiDAR LiDAR Pseudo-LiDAR Figure 21: Pseudo-LiDAR quality, set 1. Measured LiDAR against the UniK3D pseudo-LiDAR used as the pretraining target, in top-down and oblique views. Top-down ViewOblique View Input Images LiDAR Pseudo-LiDAR LiDAR Pseudo-LiDAR Figure 22: Pseudo-LiDAR quality, set 2. Front LeftFrontFront Right Ground Truth Det. Only Det. + Map Det. + Map + Occ. Figure 23: Detection pretraining quality. Each panel concatenates the front-left, front, and front-right camera views. Ground- truth and predicted 3D boxes are shown for detection-only, detection–map, and detection–map–occupancy pretraining. Front LeftFrontFront Right Ground Truth Det. Only Det. + Map Det. + Map + Occ. Figure 24: Detection pretraining quality, sample 4. Front LeftFrontFront Right Ground Truth Det. Only Det. + Map Det. + Map + Occ. Figure 25: Detection pretraining quality, sample 5. Front LeftFrontFront Right Ground Truth Det. Only Det. + Map Det. + Map + Occ. Figure 26: Detection pretraining quality, sample 6. Front LeftFrontFront Right Ground Truth Det. Only Det. + Map Det. + Map + Occ. Figure 27: Detection pretraining quality, sample 7. Input Images Ground Truth Map OnlyDet. + MapDet. + Map + Occ. Figure 28: Map pretraining quality after the shared 20-epoch budget. Columns compare map-only, detection–map, and detection– map–occupancy pretraining against ground truth. The fixed 20-epoch budget is not sufficient for the map task to converge or reach its best attainable quality, so these outputs should be interpreted as intermediate pretraining states. Input Images Ground Truth Occ. OnlyDet. + Map + Occ. Figure 29: Occupancy pretraining quality after the shared 20-epoch budget. Columns compare occupancy-only and joint detection–map–occupancy pretraining against ground truth. Ground Truth VAD-Tiny + PAVER Figure 30: Downstream detection quality on nuScenes. All detection classes are projected onto the front camera at a common score threshold. Ground Truth VAD-Tiny + PAVER Figure 31: Additional downstream detection examples on nuScenes. Input Images Ground Truth VAD-Tiny + PAVER Figure 32: Downstream planning quality. After ground truth, the columns are VAD-Tiny trained from scratch and the same model initialized with PAVER pretraining. Input Images Ground Truth VAD-Tiny + PAVER Figure 33: Joint map–planning and curved-driving cases. The first three rows improve both planning L2 and vector-map F1; the remaining rows show curved trajectories with lower planning L2. Input Images Ground Truth VAD-Tiny w/o Action Statew/ Action State Figure 34: Action-state ablation. After ground truth, the columns are VAD-Tiny trained from scratch, PAVER pretrained without action-state conditioning, and full PAVER. Input Images Ground Truth VAD-Tiny w/o Maskw/ Mask Figure 35: Action-corridor masking ablation. After ground truth, the columns are VAD-Tiny trained from scratch, PAVER pretrained without action-corridor masking, and full PAVER. Input Images Ground Truth VAD-Tiny PAVER (10K)PAVER (30K) Figure 36: Auxiliary capacity quality, PAVER (30K). After ground truth, the columns are VAD-Tiny trained from scratch and PAVER pretrained with 10K and with 30K auxiliary parameters. Input Images Ground Truth VAD-Tiny PAVER (90K) Figure 37: Auxiliary capacity quality, PAVER (90K). After ground truth, the columns are VAD-Tiny trained from scratch and PAVER pretrained with 90K auxiliary parameters. Input Images Ground Truth VAD-Tiny PAVER (Pseudo-LiDAR) Figure 38: Pretraining-target quality, pseudo-LiDAR. After ground truth, the columns are VAD-Tiny trained from scratch and PAVER pretrained with pseudo-LiDAR as the target source. Pseudo-LiDAR replaces measured LiDAR as the input to the target builder, while the sparse action-target formulation and pretraining objective remain unchanged. Input Images Ground Truth VAD-TinyDet. + Map + Occ. Figure 39: Pretraining-target quality, detection, map and occupancy. After ground truth, the columns are VAD-Tiny trained from scratch and a model pretrained on detection, vector map, and occupancy jointly. Input Images Ground TruthVAD-Base+ PAVER Figure 40: Cross-model quality, VAD-Base. After ground truth, the columns are the reproduced epoch-60 model and the same architecture initialized with PAVER pretraining. Input Images Ground TruthGenAD+ PAVER Figure 41: Cross-model quality, GenAD. After ground truth, the columns are GenAD trained from scratch and the same model initialized with PAVER pretraining. Front LeftFront Front Right Input Image VAD-Tiny + PAVER Figure 42: Grad-CAM results on the nuScenes validation set. Front LeftFront Front Right Input Image VAD-Tiny + PAVER Figure 43: Grad-CAM results on the nuScenes validation set. Front LeftFront Front Right Input Image VAD-Tiny + PAVER Figure 44: Grad-CAM results on the nuScenes validation set. Front LeftFront Front Right Input Image VAD-Tiny + PAVER Figure 45: Grad-CAM results on the nuScenes validation set. Front LeftFront Front Right Input Image VAD-Tiny + PAVER Figure 46: Grad-CAM results on the nuScenes validation set. Front LeftFront Front Right Input Image VAD-Tiny + PAVER Figure 47: Grad-CAM results on the nuScenes validation set. Front LeftFront Front Right Input Image VAD-Tiny + PAVER Figure 48: Grad-CAM response to a road-surface stop marking on the nuScenes validation set. PAVER concentrates attribution on the marking. Front LeftFront Front Right Input Image VAD-Tiny + PAVER Figure 49: Grad-CAM results on the nuScenes validation set. Front LeftFront Front Right Input Image VAD-Tiny + PAVER Figure 50: Grad-CAM results on the nuScenes validation set. Front LeftFront Front Right Input Image VAD-Tiny + PAVER Figure 51: Grad-CAM results on the nuScenes validation set. Front LeftFront Front Right Input Image VAD-Tiny + PAVER Figure 52: Grad-CAM results on the nuScenes validation set. Front LeftFront Front Right Input Image VAD-Tiny + PAVER Figure 53: Grad-CAM results on the nuScenes validation set. Front LeftFront Front Right Input Image VAD-Tiny + PAVER Figure 54: Grad-CAM results on the nuScenes validation set. Front LeftFront Front Right Input Image VAD-Tiny + PAVER Figure 55: Grad-CAM results on the nuScenes validation set. Front LeftFront Front Right Input Image VAD-Tiny + PAVER Figure 56: Grad-CAM results on the nuScenes validation set. Front LeftFront Front Right Input Image VAD-Tiny + PAVER Figure 57: Grad-CAM results on the nuScenes validation set. Front LeftFront Front Right Input Image VAD-Tiny + PAVER Figure 58: Grad-CAM results on the nuScenes validation set. Front LeftFront Front Right Input Image VAD-Tiny + PAVER Figure 59: Grad-CAM results on the nuScenes validation set. Front LeftFront Front Right Input Image VAD-Tiny + PAVER Figure 60: Grad-CAM results on the nuScenes validation set. Front LeftFront Front Right Input Image VAD-Tiny + PAVER Figure 61: Grad-CAM results on the nuScenes validation set. Front LeftFront Front Right Input Image VAD-Tiny + PAVER Figure 62: Grad-CAM results on the nuScenes validation set. Front LeftFront Front Right Input Image VAD-Tiny + PAVER Figure 63: Grad-CAM results on the nuScenes validation set. Front LeftFront Front Right Input Image Scratch (BEV) PAVER (BEV) Scratch (Plan.) PAVER (Plan.) Figure 64: Grad-CAM spatial localization for VAD-Base, scene 1. Grad-CAM maps for the BEV and planning targets with and without PAVER pretraining, shown for the three front cameras. Front LeftFront Front Right Input Image Scratch (Det.) PAVER (Det.) Scratch (Mot.) PAVER (Mot.) Scratch (Map) PAVER (Map) Figure 65: Grad-CAM spatial localization for VAD-Base, scene 1. Grad-CAM maps for the detection, motion and map targets with and without PAVER pretraining, shown for the three front cameras. Front LeftFront Front Right Input Image Scratch (BEV) PAVER (BEV) Scratch (Plan.) PAVER (Plan.) Figure 66: Grad-CAM spatial localization for VAD-Base, scene 3. Grad-CAM maps for the BEV and planning targets with and without PAVER pretraining, shown for the three front cameras. Front LeftFront Front Right Input Image Scratch (Det.) PAVER (Det.) Scratch (Mot.) PAVER (Mot.) Scratch (Map) PAVER (Map) Figure 67: Grad-CAM spatial localization for VAD-Base, scene 3. Grad-CAM maps for the detection, motion and map targets with and without PAVER pretraining, shown for the three front cameras. Front LeftFront Front Right Input Image Scratch (BEV) PAVER (BEV) Scratch (Plan.) PAVER (Plan.) Figure 68: Grad-CAM spatial localization for VAD-Base, scene 4. Grad-CAM maps for the BEV and planning targets with and without PAVER pretraining, shown for the three front cameras. Front LeftFront Front Right Input Image Scratch (Det.) PAVER (Det.) Scratch (Mot.) PAVER (Mot.) Scratch (Map) PAVER (Map) Figure 69: Grad-CAM spatial localization for VAD-Base, scene 4. Grad-CAM maps for the detection, motion and map targets with and without PAVER pretraining, shown for the three front cameras. Front LeftFront Front Right Input Image Scratch (BEV) PAVER (BEV) Scratch (Plan.) PAVER (Plan.) Figure 70: Grad-CAM spatial localization for VAD-Base, scene 5. Grad-CAM maps for the BEV and planning targets with and without PAVER pretraining, shown for the three front cameras. Front LeftFront Front Right Input Image Scratch (Det.) PAVER (Det.) Scratch (Mot.) PAVER (Mot.) Scratch (Map) PAVER (Map) Figure 71: Grad-CAM spatial localization for VAD-Base, scene 5. Grad-CAM maps for the detection, motion and map targets with and without PAVER pretraining, shown for the three front cameras.

Instant research discovery