Diffusion models have become a popular choice for decision-making tasks in robotics, and more recently, are also being considered for solving autonomous driving tasks. However, their applications and evaluations in autonomous driving remain limited to simulation-based or laboratory settings. The full strength of diffusion models for large-scale, complex real-world settings, such as End-to-End Autonomous Driving (E2E AD), remains underexplored. In this study, we conducted a systematic and large-scale investigation to unleash the potential of the diffusion models as planners for E2E AD, based on a tremendous amount of real-vehicle data and road testing. Through comprehensive and carefully controlled studies, we identify key insights into the diffusion loss space, trajectory representation, and data scaling that significantly impact E2E planning performance. Moreover, we also provide an effective reinforcement learning post-training strategy to further enhance the safety and robustness of the learned planner. The resulting diffusion-based learning framework, Hyper Diffusion Planner (HDP), is deployed on a real-vehicle platform and evaluated across 6 urban driving scenarios and 200 km of real-world testing, achieving a notable 10x performance improvement over the base model. Our work demonstrates that diffusion models, when properly designed and trained, can serve as effective and scalable E2E AD planners for complex, real-world autonomous driving tasks.
Figures & tables
Figure 1: Simulation-based benchmarks differ substantially from real-world deployment, leaving the true capability of diffusion-based planners largely unverified. Through a detailed investigation of the IL/RL training recipe, we unleash the potential of diffusion-based planning, achieving a 10x improvement in real world.
Figure 2: Model architecture of the base model.
Table 3
Figure 5: Comparison of prediction quality with waypoint and velocity representations.
Figure 6: Comparison of waypoint, velocity representations and velocity with hybrid loss, which yields both steadiness and smoothness.
Figure 6
Figure 9: Data scaling experiments. Both open- and closed-loop performance improve as data scales up.
Model Name
Data Size
Open-Loop Score
Closed-Loop Score
Success Rate
Stability Score
Overall Score
Base Model
M
51.07
15.67
0.00
7.83
with τ0 -loss & τ0 -pred
M
75.27
22.84
0.00
11.42
+ Velocity Supervision
M
84.38
34.72
9.24
21.98
+ Hybrid Loss
M
85.05
61.88
53.88
57.88
+ Data Scaling
L
86.07
70.59
59.00
64.79
Table 2: Main results. a represents the highest score. The open-loop score is evaluated through data replay on test datasets, while the closed-loop score is obtained from real-world road testing on a real-vehicle platform.
Figure 10: (a) Relative success rate vs. dataset volume in frequently-occurring scenarios; (b) stability score under the same scaling; (c) relative success rate before and after RL fine-tuning in safety-critical scenarios.
Figure 11: Closed-loop real-vehicle testing, illustrated with two representative frames.
Figure 12: Replay of bad cases before and after post-training. HDP-RL † in blue , HDP in red .
Figure 13: Reward versus training steps: HDP-RL with vs. without the RL-hybrid loss in Eq. ( 9 ). The subgraph is the image after rescaling the coordinates.
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 14: Closed-loop real-vehicle testing results. Each row contains representative frames from the scenario.
τ0 -pred.
vt -pred.
ϵ -pred.
τ0 -loss: E∥τ^0−τ0∥2
τ^0=τθ
τ^0=αtτt−σtvθ;t
τ^0=(τt−σtϵθ)/αt
vt -loss: E∥v^t−vt∥2
v^t=(αtτt−τθ)/σt
v^t=vθ;t
v^t=(ϵθ−σtτt)/αt
ϵ -loss: E∥ϵ^−ϵ∥2
ϵ^=(τt−αtτθ)/σt
ϵ^=σtτt+αtvθ;t
ϵ^=ϵθ
Appendix
Table 3: Mutual conversions among the three diffusion quantities. Columns index the model parameterization and rows index the loss space. Predicted quantities are denoted with ⋅^ and the network is parameterized by θ . Diagonal entries are the canonical objectives; off-diagonal entries first transform the network output into the target space and then compute the squared error.
Figure 15
Type
Parameter
Symbol
Value
IL
Num. block
-
6
Dim. hidden layer
-
256
Num. multi-head
-
8
Hybrid loss weight
ω
0.1
RL
Group size
-
32
Temperature
β
1.0
Appendix
Table 6: Hyperparameters of HDP / HDP-RL
Model Name
Data Size
Open-Loop Score
Closed-Loop Score
Success Rate
Stability Score
Overall Score
Regression Model
XL
73.98
21.81
0.00
10.91
Base Model
M
51.07
15.67
0.00
7.83
with τ0 -loss & τ0 -pred
M
75.27
22.84
0.00
11.42
HDP
XL
88.94
71.24
79.53
75.38
HDP-RL
-
-
83.49
84.65
84.07
Appendix
Table 7: Compare the open-loop and closed-loop scores with baselines.
Figure 16: Visualization of anchor-based truncated diffusion using (a) offset prediction and (b) direct prediction. The ground truth is in red and model predictions in purple. The deeper the trajectory is, the higher the predicted score.
Model Name
Closed-Loop Score
Success Rate
Stability Score
Overall Score
HDP-RL † (only safety reward)
72.89
79.53
76.20
DPPO-style methods (multi-rewards)
65.50
89.76
77.63
HDP-RL (multi-rewards)
83.49
84.65
84.07
Appendix
Table 8: Compare the closed-loop scores with DPPO-style RL methods ( Black et al., 2024b ; Ren et al., 2025 ) .
Diffusion-based planning generates diverse action sequences or trajectories by progressively denoising candidates conditioned on goals, scenes, and constraints.
End-to-end planning has emerged as a dominant paradigm for autonomous driving, where recent models often adopt a scoring-selection framework to choose trajectories from a large set of candidates, with diffusion-based decoding showing strong promise. However, directly selecting from the entire candidate space remains difficult to optimize, and Gaussian perturbations used in diffusion often introduce unrealistic trajectories that complicate the denoising process. In addition, for training these models, reinforcement learning (RL) has shown promise, but existing end-to-end RL approaches typically rely on a single coupled reward without structured signals, limiting optimization effectiveness. To address these challenges, we propose HAD, an end-to-end planning framework with a Hierarchical Diffusion Policy that decomposes planning into a coarse-to-fine process. To improve trajectory generation, we introduce Structure-Preserved Trajectory Expansion, which produces realistic candidates while maintaining kinematic structure. For policy learning, we develop Metric-Decoupled Policy Optimization (MDPO) to enable structured RL optimization across multiple driving objectives. Extensive experiments show that HAD achieves new state-of-the-art performance on both NAVSIM and HUGSIM, outperforming prior arts by a huge margin: +2.3 EPDMS on NAVSIM and +4.9 Route Completion on HUGSIM.
End-to-end diffusion planning has shown strong potential for autonomous driving, but the physical feasibility of generated trajectories remains insufficiently addressed. In particular, generated trajectories may exhibit local geometric irregularities, violate trajectory-level kinematic constraints, or deviate from the drivable area, indicating that the commonly used noise-centric formulation in diffusion planning is not yet well aligned with the trajectory space where feasibility is more naturally characterized. To address this issue, we propose FeaXDrive, a feasibility-aware trajectory-centric diffusion planning method for end-to-end autonomous driving. The core idea is to treat the clean trajectory as the unified object for feasibility-aware modeling throughout the diffusion process. Built on this trajectory-centric formulation, FeaXDrive integrates adaptive curvature-constrained training to improve intrinsic geometric and kinematic feasibility, drivable-area guidance within reverse diffusion sampling to enhance consistency with the drivable area, and feasibility-aware GRPO post-training to further improve planning performance while balancing trajectory-space feasibility. Experiments on the NAVSIM benchmark show that FeaXDrive achieves strong closed-loop planning performance while substantially improving trajectory-space feasibility. These findings highlight the importance of explicitly modeling trajectory-space feasibility in end-to-end diffusion planning and provide a step toward more reliable and physically grounded autonomous driving planners.
Baoyun Wang, Zhuoren Li, Ran Yu +6
College of Automotive and Energy Engineering, Tongji University, Shanghai, 201804, China · College of Transportation Engineering, Tongji University, Shanghai, 201804, China · School of Mechanical and Aerospace Engineering, Nanyang Technological University, 639798, Singapore
End-to-End (E2E) autonomous driving models have shown growing capability in recent years, with performance improving on increasingly challenging benchmarks. However, modern generative E2E planners still suffer from a substantial number of catastrophic failures in safety-critical scenarios. We find that many such failures arise from violations of physical constraints and safety requirements, leading to unsafe behavior. Motivated by this finding, in this paper, we focus on improving safety outcomes in generative end-to-end driving with a targeted reduction of catastrophic planning failures, instead of enhancing average planning quality. Towards this end, we propose DriveSafer, a failure-aware safety framework for end-to-end planners. DriveSafer explicitly steers generative planners towards safe behaviors leveraging both training-time safety constraints and inference-time safety guidance. Compared to the state-of-the-art DiffusionDrive model, on the NAVSIM benchmark, DriveSafer reduces the number of catastrophic failures (PDMS=0) by 48%, with over 65% reduction in drivable-area compliance failures.