MeanFuser: Fast One-Step Multi-Modal Trajectory Generation and Adaptive Reconstruction via MeanFlow for End-to-End Autonomous Driving
Authors: Junli Wang, Yinan Zheng, Xueyi Liu, Zebin Xing, Pengfei Li, Guang Li, Kun Ma, Guang Chen, Hangjun Ye, Zhongpu Xia, Long Chen, Qichao Zhang
Organizations: SKL-MAIS, Institute of Automation, Chinese Academy of Sciences · School of Artificial Intelligence, University of Chinese Academy of Sciences · Xiaomi EV · Institute for AI Industry Research (AIR), Tsinghua University
Generative models have shown great potential in trajectory planning. Recent studies demonstrate that anchor-guided generative models are effective in modeling the uncertainty of driving behaviors and improving overall performance. However, these methods rely on discrete anchor vocabularies that must sufficiently cover the trajectory distribution during testing to ensure robustness, inducing an inherent trade-off between vocabulary size and model performance. To overcome this limitation, we propose MeanFuser, an end-to-end autonomous driving method that enhances both efficiency and robustness through three key designs. (1) We introduce Gaussian Mixture Noise (GMN) to guide generative sampling, enabling a continuous representation of the trajectory space and eliminating the dependency on discrete anchor vocabularies. (2) We adapt ``MeanFlow Identity" to end-to-end planning, which models the mean velocity field between GMN and trajectory distribution instead of the instantaneous velocity field used in vanilla flow matching methods, effectively eliminating numerical errors from ODE solvers and significantly accelerating inference. (3) We design a lightweight Adaptive Reconstruction Module (ARM) that enables the model to implicitly select from all sampled proposals or reconstruct a new trajectory when none is satisfactory via attention weights.Experiments on the NAVSIM closed-loop benchmark demonstrate that MeanFuser achieves outstanding performance without the supervision of the PDM Score and exceptional inference efficiency, offering a robust and efficient solution for end-to-end autonomous driving. Our code and model are available at https://github.com/wjl2244/MeanFuser.
Figures & tables
Figure 1: (a) illustrates the differences between our proposed method and existing generative approaches, highlighting the introduction of Gaussian mixture noise to replace anchor vocabularies, one-step sampling, and the adaptive reconstruction module. (b) shows the advantages of MeanFuser over GoalFlow [ 33 ] , Hydra-MDP [ 20 ] , and DiffusionDrive [ 22 ] in terms of closed-loop performance, inference speed and plan module inference speed.
Figure 2: Failure scene visualization. Anchor-guided models (GoalFlow, DiffusionDrive) fail due to the inability of discrete vocabularies to cover the entire trajectory space, while our model generates proposals that encompass the optimal trajectories.
Figure 3: Overall architecture of MeanFuser. Training : During training, both the images and ego-vehicle states are encoded into context features, with auxiliary supervision from mapping tasks. The model is conditioned on these context features to learn the average velocity field uθ over the time interval r and t . Multi-Modal Sample : Noise samples are drawn from Gaussian Mixture Noise, and the one-step sampling formulation is then applied to generate diverse multi-modal trajectories. Adaptive Reconstruction Module : The sampled multi-modal trajectories are encoded and fused with the context features through cross-attention, after which a Projector outputs the final planning trajectory.
Method
Venue
Input
Img. Backbone
NC ↑
DAC ↑
TTC ↑
Comf. ↑
EP ↑
PDMS ↑
TransFuser [ 7 ]
IEEE TPAMI
C & L
ResNet-34
97.7
92.8
92.8
100
79.2
84.0
VADv2 [ 3 ]
arXiv 2024
C & L
ResNet-34
97.2
89.1
91.6
100
76.0
80.9
Hydra-MDP [ 20 ]
arXiv 2024
C & L
ResNet-34
98.3
96.0
94.6
100
78.7
86.5
GoalFlow* [ 33 ]
CVPR 2025
C & L
ResNet-34
98.3
93.8
94.3
100
79.8
85.7
DiffusionDrive [ 22 ]
CVPR 2025
C & L
ResNet-34
98.2
96.2
94.7
100
82.2
88.1
WoTE [ 19 ]
ICCV 2025
C & L
ResNet-34
98.5
96.8
94.9
99.9
81.9
88.3
Table 1: Performance on the NAVSIMv1 navtest benchmark. “C” denotes camera, and “L” denotes LiDAR. * indicates results reported from the official papers.
Method
NC ↑
DAC ↑
DDC ↑
TLC ↑
EP ↑
TTC ↑
LK ↑
HC ↑
EC ↑
EPDMS ↑
Ego Status MLP [ 21 ]
93.1
77.9
92.7
99.6
86.0
91.5
89.4
98.3
85.4
64.0
TransFuser [ 7 ]
96.9
89.9
97.8
99.7
87.1
95.4
92.7
98.3
87.2
76.7
Hydra-MDP++ [ 18 ]
97.2
97.5
99.4
99.6
83.1
96.5
94.4
98.2
70.9
81.4
DriveSuprim [ 36 ]
97.5
96.5
99.4
99.6
88.4
96.6
95.5
98.3
77.0
83.1
DiffusionDrive † [ 22 ]
98.2
96.3
99.4
99.8
87.4
97.4
97.0
98.3
87.7
88.3
MeanFuser(Ours)
98.3
97.2
99.6
99.8
87.6
97.4
97.3
98.3
88.2
89.5
Table 2: Performance of the NAVSIMv2 navtest benchmark. † denotes testing with the official checkpoint
Method
Dim
Params (M)
PDMS ↑
Plan FPS ↑
FPS ↑
TransFuser
256
55.9
84.0
3934
63
GoalFlow
256
62.3
85.7
11
10
Hydra-MDP
256
65.2
86.5
25
20
DiffusionDrive
256
60.7
88.1
75
39
MeanFuser
128
54.6
89.0
434
59
Table 3: Model parameter size, inference speed, and performance. Bold and underlined values denote the best and second-best results, respectively. “Dim” indicates the number of hidden neurons in the model, “FPS” represents the median inference speed measured on a single NVIDIA H20 GPU over multiple runs, and “Plan FPS” refers to the inference speed of trajectory planning excluding the perception encoder.
Model
Components
PDMS ↑
Nproposals
PL2>0.2↓
PL2>0.5↓
NDAC=0↓
-
DiffusionDrive
88.1
20
72.7%
20.0%
84
Base
TransFuser
84.0
-
-
-
-
M0
Base + vanilla MeanFlow
87.3 +3.3
16
92.8%
40.6%
143
M1
M0 + Guassian-mixed Noise ( GMN )
88.2 +0.9
16
69.3%
18.5%
58
M2
M1 + Adpative Reconstruction Module ( ARM )
89.0 +0.8
17
67.1%
16.9%
48
M3
M1 + Trajectory Proposals Averaging
71.2 -17.8
17
69.1%
18.0%
57
Table 4: Ablation study on the impact of each module. Base denotes the TransFuser [ 7 ] baseline. M0 replaces the decoder MLP in TransFuser with a MeanFlow model and employs a classifier to score the parallel sampled trajectory proposals. M1 and M2 further enhance M0 by introducing GMN and an ARM, respectively. M3 employs an averaging strategy across all sampled traiectories. Nproposals denotes the number of sampled trajectory proposals; PL2>0.2 and PL2>0.5 represent the proportions of scenarios where all proposals have a minimum L2 distance greater than 0.2 and 0.5 from the expert trajectory; and NDAC=0 indicates the number of cases where all proposals leave the drivable area.
Figure 4: Visualization of sampling from different Gaussian components. Parallel sampling of trajectories from distinct Gaussian components can generate diverse driving styles, ranging from conservative to aggressive.
Figure 5: Visualization of the multimodal trajectory of the model. The left image shows expert demonstration trajectories, while the right image displays the model’s inferred strategies for maintaining a straight path and performing a left lane change, representing two different modes.
Method
RC ↑
IS ↑
DS ↑
Latent TransFuser [ 7 ]
95.18 ±0.45
0.38 ±0.05
37.31 ±5.35
TransFuser [ 7 ]
93.38 ±1.20
0.50 ±0.06
47.30 ±5.72
DiffusionDrive [ 22 ]
94.16 ±1.46
0.69 ±0.02
64.27 ±2.43
MeanFuser (Ours)
94.65 ±1.32
0.73 ±0.05
70.08 ±3.20
Table 5: Longest6 Benchmark Results. We show the mean and std for all metrics (RC: Route Completion, IS: Infraction Score. DS: Driving Score).
Figure 6: Visualization on the CARLA Longest6 benchmark. The two images in the upper row and the two in the lower row showcase the planning outcomes, depicting daytime and nighttime conditions, respectively.
Figure 7: Visualization of alternative approaches for generating Gaussian Mixture Noise (GMN) . (a) Mean and standard deviation are derived from clustered expert demonstrations in the training set. (b) Mean and standard deviation are obtained through manually design.
Ngaussian
NC ↑
DAC ↑
TTC ↑
Comf. ↑
EP ↑
PDMS ↑
2
98.1
96.9
94.0
100
81.8
88.4 -0.5
8
98.6
97.0
95.0
100
82.8
89.0
16
98.1
97.3
93.8
99.9
83.3
88.8 -0.2
32
98.0
97.0
94.3
99.9
82.8
88.5 -0.5
Table 6: Number of Gaussian components and model performance. Ngaussian denotes the number of Gaussian components in the Gaussian Mixture Noise distribution.
Method
NC ↑
DAC ↑
TTC ↑
Comf. ↑
EP ↑
PDMS ↑
Data Clustering
98.6
97.0
95.0
100
82.8
89.0
Manual Design
98.2
97.0
94.1
100
83.0
88.6 -0.4
Table 7: Comparison of the GMN Generation Methods.
Method
GMN
K
PDMS ↑
D↑
MDP↑
TransFuser
-
8
94.0
0.0
-
MeanFuser
×
8
88.3
0.25
22.07
MeanFuser
√
8
89.0
0.30
26.70 +20.84%
Table 8: Comparison of multimodality and performance. (GMN: Gaussian Mixture Noise. K : number of multimodal trajectories; D : multimodality metric.)
Figure 8: Visualization of the intersection and union dynamics of bounding boxes (Car Bbox) for ego trajectories across different timesteps.
End-to-end autonomous driving maps current observations directly to future trajectories, yet those trajectories must remain valid as the scene evolves. Future state modeling aims to address this temporal mismatch, but general representations often contain information unrelated to ego planning and affect trajectory generation only through auxiliary supervision, static conditioning, or proposal evaluation. We propose FeasibleFlow, a one-step end-to-end generative framework that jointly transports a configuration-space feasibility field and multimodal ego trajectories. Our Asymmetric Joint MeanFlow uses the pathwise Jacobian-vector product in the MeanFlow identity to incorporate field evolution into trajectory transport. Because safety feedback is sparser than progress feedback, we further introduce the Anchor-relative ranker (ARR) and Pareto-ReinFlow to balance safety and progress in candidate selection and generation, respectively. Experiments on the NAVSIM benchmark demonstrate the strong performance of FeasibleFlow and validate both the joint transport of feasibility and trajectories and the proposed safety-first mechanisms.
End-to-end autonomous driving must balance multimodal maneuver generation against real-time inference constraints. Diffusion planners capture diverse behaviors, but their iterative denoising incurs considerable inference cost in safety-critical deployment, while one-step alternatives still imitate only the single expert demonstration, so neither family certifies both the diversity and the quality of its outputs. To address this, we propose Vault, a framework that couples one-step latent generation with sample-based reinforcement learning guidance. Built on a V-JEPA 2.1 visual representation pretrained on the full NAVSIM training imagery, Vault generates trajectory candidates in a single forward pass by drifting in a VAE latent space. During training, a reward-gated positive pool collects the maximum-scoring trajectories that the model itself discovers under the official evaluator and anchors the drift targets to these certified samples. This closes a reinforcement-learning loop purely through sample-based target construction, without policy gradients, learned reward models, or extra interaction: the generator imitates the frontier of its own certified successes, so quality and multimodality improve together under a single fixed sampling configuration. A learned scorer that predicts the official score and its sub-metrics then selects the executed trajectory, and all training-only modules are discarded at inference. On the NAVSIM benchmarks, Vault achieves state-of-the-art performance at real-time inference cost, reaching 94.6 PDMS on v1 and 91.2 EPDMS on v2.
End-to-end planning in complex scenes must model diverse driving intentions and select the most appropriate trajectory using scene context and future risks. However, multi-query regression may collapse to one mode, while anchor-based and generative planners share a scaling bottleneck: small candidate sets limit maneuver coverage or discovery of high-quality trajectories, whereas large sets increase generation and scoring costs. We propose ManeuverFlow to generate trajectories for distinct intentions in one step and evaluate and refine them using current and future evidence. The model selects relevant maneuver modes, draws corresponding noise from the Gaussian mixture prior, and maps it to intention-specific trajectories through a maneuver-routed mixture-of-experts MeanFlow field. Current features and future features predicted by the world model are aggregated along each candidate path into trajectory-specific evidence for PDM-factor prediction. Using the same evidence, the refiner predicts local Frenet residuals for selected candidates; refined and retained candidates are then jointly re-scored for selection. With a lightweight TransFuser-style ResNet-34 backbone, ManeuverFlow achieves 90.6 PDMS and 90.1 EPDMS on NAVSIM v1/v2 navtest, respectively. On WorldEngine's recent 3DGS closed-loop benchmark, it attains the highest listed PDMS∗ scores of 75.19 and 74.58 under logged and IDM-controlled traffic, respectively. Code: https://github.com/wyh-autonomy/maneuverflow.
Yuhao Wei, Yong Chen, Chen Yang +7
Chongqing University · Artificial Intelligence Center, Geely Automotive Research Institute