AD-E2E-JEPA: A Joint-Embedding Predictive Architecture For End-to-End Autonomous Driving

Abstract: Autonomous driving requires \textit{world models} that can understand the physical world, reason and plan, and operate safely. In this paper, we first systematically evaluate existing action-conditioned joint-embedding predictive architecture (JEPA) world models, including LeWM, DINO-WM, and JEPA-WM for end-to-end autonomous driving (E2EAD). To isolate world-model quality from policy learning, we employ a goal-conditioned zero-shot planning setting that evaluates these models using ground-truth future observations as goals, without training any driving policy. We find that existing JEPA-based world models are either accurate for driving but computationally expensive, or computationally efficient but insufficient for planning. To address this trade-off, we propose \textbf{AD-E2E-JEPA}, which introduces a SIGReg-regularized learnable projector applied to projected patch embeddings. The projector reduces the number of planning patches by $16\times$ and the embedding dimension by $4\times$, achieving a $100\times$ inference speedup while retaining planning performance, with a 0.8-second runtime for an 8-frame rollout over 256 candidate trajectories. \textit{Without} training any driving policy, the world model itself reaches the goals located 20 meters away on average within the displacement of respectively 4.0/2.8 meters, using world-model rollouts over trajectory vocabularies of respectively 256/8,192 candidates. On the NAVSIMv2 benchmark, it achieves 67.3/72.9 EPDMS with multiplicative safety metrics and 84.1/86.5 EPDMS$<sup>{\dagger}$ without them in goal-conditioned zero-shot planning. Experiments further show that the self-supervised pretrained projector improves downstream imitation learning performance from 80.2 to 85.4 EPDMS. The source code is available at https://github.com/HaoranZhuExplorer/AD-E2E-JEPA

• The paper proposes a SIGReg-regularized spatial and channel projector compressing dense visual embeddings to enable zero-shot planning in an autonomous driving world model
• It introduces a unique approach to predicting future visual representations under ego trajectories, shortening planning time to approximately 0.8 seconds with competitive performance on the NAVSIM benchmark
• The study demonstrates significant improvements in driving metrics, including EPDMS, geodesic metrics, and hit rates, and shows the benefit of rollout training and data scaling

AD-E2E-JEPA frames end-to-end autonomous driving as a latent-dynamics planning problem rather than exclusively as reactive imitation learning. Its central premise is that an action-conditioned JEPA world model can predict future visual representations under candidate ego trajectories, allowing an autonomous vehicle to select actions by comparing predicted future states with a specified visual goal. The paper’s principal contribution is an efficiency-oriented modification of JEPA-based world modeling: a learnable spatial and channel projector, regularized with SIGReg, compresses dense visual embeddings sufficiently to make zero-shot planning practical while preserving the planning-relevant information of the original representation (2609.34085).

Problem formulation and research objectives

The paper focuses on front-camera-only E2EAD. At each time step, the model receives a temporal context of images and ego poses and predicts future ego poses. Rather than training a policy directly to imitate the demonstrated future trajectory, the proposed system uses the demonstrated ego-motion sequence as action conditioning during world-model training. The distinction is important: human trajectories provide supervision for learning environment dynamics, but the final zero-shot planning experiment does not train a driving policy or trajectory scorer.

The planning task is goal-conditioned. A future observation, specifically an image several frames ahead, is treated as the goal. The model rolls out multiple candidate trajectories and selects the trajectory whose predicted future latent representation is closest to the latent representation of the goal image. This isolates the quality of the learned world model from the quality of a policy trained on expert demonstrations. The evaluation therefore addresses a specific question: whether a JEPA model trained on offline driving data can rank candidate trajectories according to their ability to reach a visual target.

The authors first establish a computational and performance trade-off among existing JEPA-based world models. LeWM uses a global representation and is fast, but its planning accuracy is poor. DINO-WM and JEPA-WM retain dense patch embeddings and provide substantially better planning reliability, but their inference cost is incompatible with efficient autonomous-driving planning. AD-E2E-JEPA is designed to occupy the intermediate regime: it retains dense spatial information in compressed form and uses a trajectory vocabulary instead of iterative CEM optimization.

AD-E2E-JEPA architecture

The baseline world model uses a frozen DINOv3 ViT-L encoder, an AdaLN-style predictor with RoPE, and a linear action encoder. The action at time $t$ is the relative ego-pose transformation between consecutive frames:

$$\mathbf{a}_t = [\Delta x_{t\rightarrow t+1}, \Delta y_{t\rightarrow t+1}, \Delta\theta_{t\rightarrow t+1}]^\top.$$

The predictor receives historical visual embeddings and action embeddings and estimates the embeddings of subsequent frames. In the baseline formulation, prediction is performed directly in the DINOv3 embedding space using MSE. Because the target encoder is frozen, the baseline does not require an additional anti-collapse objective.

AD-E2E-JEPA inserts a learnable projector between the frozen DINOv3 encoder and the action-conditioned predictor. The projector consists of two convolutional layers with stride $2 \times 2$. It reduces the spatial number of patch embeddings by $16\times$ and reduces the channel dimension from 1024 to 256, yielding a nominal compression of the planning representation by approximately $64\times$ in the number of scalar embedding values. The paper attributes the overall planning acceleration primarily to this reduction in the number and dimensionality of tokens.

Compression introduces a collapse risk: distinct DINOv3 embeddings could be mapped to nearly constant projected vectors. The paper addresses this with stop-gradient targets and SIGReg. SIGReg is applied independently at each projected patch location and time step across the batch, using random one-dimensional projections and an Epps–Pulley normality objective. This differs from formulations that regularize global CLS tokens or aggregate temporal statistics without preserving patchwise structure. The design assumption is that an approximately isotropic Gaussian distribution at each spatial-temporal location maintains sufficient information for prediction while preventing degenerate representations.

The resulting objective combines projected-space prediction loss with SIGReg. An optional rollout objective extends supervision across the full eight-frame future horizon. Teacher-forced predictions are supplemented by autoregressive predictions, with truncated backpropagation through time and stop-gradient through earlier rollout contexts. This explicitly penalizes compounding prediction errors, which is particularly relevant because the planning procedure evaluates multi-step world-model rollouts rather than only one-step predictions.

Zero-shot planning procedure

The planning system uses a fixed vocabulary of up to 8192 clustered driving trajectories. These trajectories are sorted by angular coordinate and can be subsampled to 256, 512, 1024, 2048, or 4096 candidates. For each candidate, the world model predicts the future latent representation under the corresponding action sequence. The selected trajectory minimizes squared latent distance to the latent embedding of the goal image.

This vocabulary-based search replaces CEM, whose iterative evaluation of large candidate populations is considered too expensive for the target setting. The resulting procedure is simple and parallelizable: all candidate rollouts can be evaluated simultaneously on an A100 GPU. It also exposes a direct trade-off between search granularity and inference time. Increasing the vocabulary from 256 to 8192 improves final-pose accuracy and EPDMS, but linearly increases planning latency.

The evaluation reports three classes of metrics. EPDMS measures NAVSIMv2 driving quality and safety, while EPDMS$^\dagger$ excludes multiplicative safety terms and therefore reflects the quality component more directly. Geodesic metrics measure final displacement, longitudinal and lateral position errors, and heading error. Top-1 and top-5 hit rates measure whether the ground-truth trajectory is ranked among the lowest-cost candidates. The latter metrics are especially informative for diagnosing the world model itself, because they assess whether latent distance assigns low cost to the demonstrated trajectory even when the selected trajectory is not exactly identical to it.

Comparative planning results

The comparison with prior world models supports the paper’s main efficiency claim. On 100 subsampled NAVSIM test scenes, LeWM requires only 0.7 seconds per scene, but obtains an EPDMS of 48.3, an FDE of 12.4 meters, and Top-1/Top-5 hit rates of 6%/18%. Dense-feature models are much more accurate: DINO-WM reaches 68.3 EPDMS and JEPA-WM reaches 74.2, with Top-1/Top-5 hit rates of 40%/73% and 45%/75%, respectively. However, their planning times are 91.8 and 101.0 seconds per scene.

AD-E2E-JEPA reaches 76.6 EPDMS with 0.8 seconds of planning time in the same 100-scene setting. Its rollout-trained variant reaches 70.4 EPDMS, with improved FDE from 4.2 to 3.5 meters and improved Top-1/Top-5 hit rates from 27%/59% to 34%/67%. The apparent reduction in EPDMS despite better geodesic and reliability metrics is one of the paper’s important non-monotonic results. It indicates that the latent-distance objective is not fully aligned with the multiplicative safety structure of NAVSIMv2. A world model can become better at identifying the ground-truth future and still produce lower safety-weighted performance on a small evaluation subset.

The strongest evidence appears on the full 12,146-scene test set. The trainval-trained rollout variant obtains an EPDMS of 67.3 and EPDMS$^\dagger$ of 84.1, with an FDE of 4.0 meters and Top-1/Top-5 hit rates of 53.8%/82.7%. In the same broad setting, LeWM obtains 39.8 EPDMS, 66.7 EPDMS$^\dagger$, 14.6-meter FDE, and 5.7%/11.3% hit rates. Thus, the proposed compressed representation substantially improves both planning accuracy and reliability relative to the efficient baseline.

The candidate-vocabulary ablation shows a consistent accuracy-latency trade-off:

Candidate trajectories	Planning time	EPDMS	EPDMS$^\dagger$	FDE	Heading error
256	0.8 s	67.3	84.1	4.0 m	3.5°
512	1.4 s	69.2	85.0	3.6 m	2.9°
1,024	2.5 s	70.5	85.5	3.2 m	2.5°
2,048	4.7 s	71.5	86.0	3.0 m	2.2°
4,096	9.3 s	72.1	86.3	2.9 m	2.1°
8,192	18.2 s	72.9	86.5	2.8 m	2.0°

The 8192-trajectory model achieves the paper’s best reported planning result: 72.9 EPDMS, 86.5 EPDMS$^\dagger$, 2.8-meter FDE, and 2.0-degree mean heading error. The implication is that AD-E2E-JEPA is not merely fast at a fixed accuracy level; it provides a tunable compute-performance curve. At 256 candidates, it performs an eight-frame rollout in approximately 0.8 seconds, compared with roughly 92–101 seconds for DINO-WM and JEPA-WM. The paper characterizes this as a $$\mathbf{a}_t = [\Delta x_{t\rightarrow t+1}, \Delta y_{t\rightarrow t+1}, \Delta\theta_{t\rightarrow t+1}]^\top.$$0 speedup, although the exact ratio relative to the reported dense-model runtimes is closer to $$\mathbf{a}_t = [\Delta x_{t\rightarrow t+1}, \Delta y_{t\rightarrow t+1}, \Delta\theta_{t\rightarrow t+1}]^\top.$$1–$$\mathbf{a}_t = [\Delta x_{t\rightarrow t+1}, \Delta y_{t\rightarrow t+1}, \Delta\theta_{t\rightarrow t+1}]^\top.$$2.

The reliability results require more careful interpretation. Although increasing the number of candidates improves EPDMS and final-pose accuracy, the reported Top-1 and Top-5 hit rates decrease from 53.8%/82.7% at 256 candidates to 15.3%/33.8% at 8192 candidates. This is not necessarily a contradiction: the hit-rate definition requires the ground-truth trajectory to rank among the top $$\mathbf{a}_t = [\Delta x_{t\rightarrow t+1}, \Delta y_{t\rightarrow t+1}, \Delta\theta_{t\rightarrow t+1}]^\top.$$3 candidates in a larger set, making the ranking problem harder as more alternatives are added. The result suggests that aggregate driving metrics and exact candidate-ranking reliability measure different properties of the world model.

Role of rollout training and data scale

Rollout training is a major component of the final system. Without rollout loss, AD-E2E-JEPA trained on trainval obtains 63.2 EPDMS, 80.1 EPDMS$$\mathbf{a}_t = [\Delta x_{t\rightarrow t+1}, \Delta y_{t\rightarrow t+1}, \Delta\theta_{t\rightarrow t+1}]^\top.$$4, and 6.2-meter FDE on the full test set. Adding rollout training increases performance to 67.3 EPDMS and 84.1 EPDMS$$\mathbf{a}_t = [\Delta x_{t\rightarrow t+1}, \Delta y_{t\rightarrow t+1}, \Delta\theta_{t\rightarrow t+1}]^\top.$$5, while reducing FDE to 4.0 meters and increasing the Top-1/Top-5 hit rate to 53.8%/82.7%. The improvement supports the claim that one-step latent prediction is insufficient for planning over a four-second horizon.

The authors also scale training data from the 10-hour navtrain split to the 70-hour trainval training portion. The benefit is clearest in combination with rollout training and full-scene evaluation, where the larger configuration attains the strongest overall result. However, the paper does not isolate the effects of data scale, projector design, rollout training, and hyperparameter changes through a complete factorial ablation. The reported trainval configuration also uses different batch sizes and SIGReg weights from the navtrain configuration. Consequently, the performance gain cannot be attributed exclusively to the additional driving data.

Transfer to imitation learning

The pretrained projector is evaluated separately as a representation for conventional imitation learning. The authors discard the world-model predictor, retain the DINOv3 encoder and learned projector, and attach a trajectory decoder based on temporal and spatial positional embeddings, command conditioning, and cross-attention. The entire model is fine-tuned using trajectory MSE.

The pretrained projector improves the reported EPDMS from 80.2 with a randomly initialized projector to 85.4, a 5.2-point gain. The component-wise NAVSIM metrics also improve, including no-collision, drivable-area compliance, driving-direction compliance, time-to-collision, lane keeping, history comfort, and extended comfort. In the benchmark comparison, the pretrained-projector model obtains 97.7 NC, 93.7 DAC, 99.2 DDC, 99.8 TLC, 87.3 EP, 96.8 TTC, 97.0 LK, 98.4 HC, and 88.7 EC, compared with 96.8, 89.8, 98.3, 99.7, 87.1, 95.7, 94.8, 98.3, and 84.1 for the random-projector variant.

This transfer result is conceptually significant because it isolates a component often obscured in end-to-end systems. The projector is not trained with trajectory labels, yet its world-model objective produces a representation that improves downstream supervised driving. The result supports the narrower claim that predictive latent compression can encode task-relevant structure beyond the immediate zero-shot planning task. It does not establish that the projector alone is sufficient for robust driving, since the encoder is DINOv3, the complete model is fine-tuned, and the evaluation remains within the NAVSIM benchmark.

Limitations and open questions

The experimental setting deliberately simplifies autonomous driving. It uses only a front camera, four historical frames, fixed candidate trajectories, and a future image as the goal. The world model therefore does not address multi-camera fusion, LiDAR, explicit map conditioning, dynamic-agent interaction, closed-loop execution, or uncertainty-aware planning. NAVSIM is a pseudo-simulation benchmark rather than a deployment evaluation, so the reported EPDMS values should not be interpreted as evidence of real-world safety.

The goal-conditioned formulation also supplies privileged information: the planner is given the ground-truth future image. In ordinary autonomous driving, a future visual goal is not generally available at planning time. The experiment is valuable for isolating world-model planning capacity, but it does not demonstrate autonomous destination selection or ordinary reactive driving. The safety metrics are likewise not optimized by the latent-distance objective. The paper explicitly reports EPDMS$$\mathbf{a}_t = [\Delta x_{t\rightarrow t+1}, \Delta y_{t\rightarrow t+1}, \Delta\theta_{t\rightarrow t+1}]^\top.$$6 because zero-shot planning does not directly enforce collision avoidance, drivable-area compliance, traffic-light compliance, or other safety constraints.

Several methodological comparisons remain incomplete. DINO-WM and JEPA-WM are evaluated on only 100 scenes because dense planning is prohibitively expensive, whereas AD-E2E-JEPA and LeWM are evaluated on all 12,146 scenes. This makes the full-test comparison against dense baselines unavailable. In addition, the 100-scene subset produces substantial variance, as reflected by differences between subset and full-test rankings. The paper also does not provide a complete ablation separating spatial compression, channel compression, SIGReg, stop-gradient, predictor architecture, rollout training, and trajectory-vocabulary search. The claim that SIGReg preserves planning information is supported by the aggregate results but not by a systematic collapse or information-retention analysis.

The relationship between latent distance and safety remains open. Increasing candidate density improves geometric accuracy and EPDMS, but decreases hit rates under the paper’s ranking definition. This suggests that the latent metric may support approximate trajectory selection without inducing a stable ordering over increasingly fine candidate sets. Whether alternative distance functions, calibrated uncertainty, safety-aware latent objectives, or hierarchical planning can resolve this discrepancy is not established.

Conclusion

AD-E2E-JEPA presents an action-conditioned JEPA world model for E2EAD in which a SIGReg-regularized projector compresses dense DINOv3 patch embeddings before latent rollout prediction. The design addresses the central computational weakness of dense JEPA planning: DINO-WM and JEPA-WM achieve strong planning quality but require approximately 1–2 minutes per scene, whereas AD-E2E-JEPA performs an eight-frame, 256-candidate rollout in approximately 0.8 seconds while retaining competitive accuracy. On 12,146 NAVSIM test scenes, the rollout-trained model reaches 67.3 EPDMS and 84.1 EPDMS$$\mathbf{a}_t = [\Delta x_{t\rightarrow t+1}, \Delta y_{t\rightarrow t+1}, \Delta\theta_{t\rightarrow t+1}]^\top.$$7; with 8192 candidates, it reaches 72.9 and 86.5, respectively, with 2.8-meter displacement and 2.0-degree heading error. The pretrained projector also improves imitation-learning performance from 80.2 to 85.4 EPDMS. The results support compressed latent world modeling as a practical planning mechanism, while leaving unresolved the transfer from privileged goal-conditioned evaluation to closed-loop, safety-constrained autonomous driving.