Source-linked AI summary
MeanFuser: Fast One-Step Multi-Modal Trajectory Generation and Adaptive Reconstruction via MeanFlow for End-to-End Autonomous Driving
Junli Wang, Yinan Zheng, Xueyi Liu, Zebin Xing, Pengfei Li, Guang Li, Kun Ma, Guang Chen, Hangjun Ye, Zhongpu Xia, Long Chen, Qichao Zhang
TL;DR
Discrete trajectory vocabularies create a coverage, robustness, and efficiency challenge for multimodal autonomous-driving planning. MeanFuser replaces that dependency with Gaussian Mixture Noise, MeanFlow-based one-step generation, and adaptive reconstruction, achieving strong NAVSIM results with fast inference and expert-trajectory-only training. Its Gaussian mixture currently fixes the mixing coefficients, leaving parameter modeling as a future direction.
Problem
Fixed trajectory vocabularies trade efficiency against robustness because large vocabularies slow inference while small ones fail under distribution shifts.
Method
MeanFuser combines continuous Gaussian Mixture Noise, MeanFlow mean-velocity modeling for one-step sampling, and an attention-based Adaptive Reconstruction Module.
Results
MeanFuser achieves 89.0 PDMS on NAVSIMv1, 89.5 EPDMS on NAVSIMv2, and 59 FPS, surpassing previous methods and running faster than several baselines.
Takeaways & Limitations
MeanFuser provides a rule-free, image-only end-to-end planner that balances one-step sampling with planning performance and avoids benchmark-score supervision.
Takeaways & Limitations
The Gaussian mixture uses fixed mixing coefficients π_k = 1, while modeling more optimal parameters remains future work.
Abstract
from arXiv · showhide
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.
1. Introduction
MeanFuser addresses the coverage and efficiency limits of discrete-anchor and multi-step generative planning with continuous Gaussian mixture noise, MeanFlow-based one-step sampling, and adaptive trajectory reconstruction. On NAVSIM, it reports strong planning performance and faster inference without benchmark-score supervision.
- Fixed trajectory vocabularies trade inference efficiency against robustness because large vocabularies slow inference while small ones fail under distribution shifts.
- MeanFuser uses Gaussian Mixture Noise to represent a broader continuous trajectory space and capture distinct driving patterns without a fixed discrete vocabulary.
- MeanFlow Identity models the mean velocity field between noise and trajectories instead of the instantaneous field used in vanilla flow matching.This design targets numerical errors from ODE solvers and supports one-step sampling.
- Adaptive Reconstruction Module attention weights select an existing proposal or reconstruct a new trajectory when sampled proposals are suboptimal.ARM relies on expert trajectories rather than benchmark evaluation submetrics.
- 89.0 PDMS on NAVSIMv1 and 89.5 EPDMS on NAVSIMv2 are achieved at 59 FPS, with inference 5.20×, 2.65×, and 1.55× faster than GoalFlow, Hydra-MDP, and DiffusionDrive.
- MeanFuser is presented as a rule-free, image-only framework that outperforms existing imitation-learning approaches on the NAVSIM closed-loop benchmark.
2. Related Work
Related work progresses from multimodal end-to-end driving and trajectory vocabularies toward diffusion and flow-based generation. These generative methods improve trajectory diversity but commonly require candidate selection and can retain discretization or sampling limitations.
- Earlier end-to-end systems fuse sensor modalities or structure planning through transformer pipelines, while VAD formulates planning as selecting trajectories from a learned vocabulary.
- Diffusion and flow-matching methods model multimodal driving behavior by injecting noise and denoising to sample diverse plausible trajectories.
- Candidate trajectory systems commonly apply rule-based or learning-based selection after generating proposals.Examples include handcrafted, map-based, or other evaluation-based scoring mechanisms.
- MeanFuser’s Adaptive Reconstruction Module decides whether to select a candidate or regenerate a trajectory based on the proposals.
3. Preliminary
The preliminary section contrasts standard flow matching with MeanFlow. Standard methods learn instantaneous velocity fields and may require many ODE-solver evaluations, whereas MeanFlow learns mean velocity to support straighter trajectories and one-step generation.
- 3.1. Problem Formulation: End-to-end planning maps raw sensor data and ego state to a future trajectory of positions and headings over a planning horizon.Training uses expert demonstration trajectories.
- 3.2. Flow-Based Model: Standard flow matching constructs a probability path from a simple prior such as Gaussian noise to a complex data distribution by learning an instantaneous velocity field.
- 3.2. Flow-Based Model: Even with a linear marginal probability path, the learned instantaneous field may produce curved individual trajectories, requiring small ODE-solver steps and multiple function evaluations.
- 3.2. Flow-Based Model: MeanFlow defines mean velocity as total displacement between times t and r divided by their interval, replacing instantaneous-velocity modeling.
- 3.2. Flow-Based Model: The MeanFlow Identity relates instantaneous and mean velocity fields and motivates training objectives that encourage straight sample trajectories.
- 3.2. Flow-Based Model: MeanFlow can generate a sample in one step using the straight-line update x_1 = x_0 + 1 · u_θ(x_0, 0, 1).
4. Method
MeanFuser encodes scene context, samples continuous multi-modal trajectories from Gaussian Mixture Noise, and uses adaptive reconstruction to produce the final plan. Its training combines mean-velocity prediction, mapping supervision, and expert-trajectory reconstruction.
- Overall pipeline: MeanFuser comprises scene-context encoding, multi-modal trajectory sampling, and adaptive reconstruction modules.Images and ego-vehicle states are encoded into context features, which condition subsequent trajectory generation and reconstruction.
- Scene Context Encoder: The scene encoder extracts image and ego-vehicle features, while auxiliary lane-map supervision supports semantic learning and convergence.The resulting scene features serve as conditional inputs for multi-modal trajectory sampling.
- Gaussian Mixture Noise: Gaussian Mixture Noise replaces standard Gaussian noise as the prior, providing continuous multi-modal trajectory guidance without discrete vocabularies.Normalized expert trajectories are clustered with k-means, and cluster statistics parameterize the Gaussian mixture components.
- Gaussian Mixture Noise: Each Gaussian component is parameterized by its mixing coefficient, mean, and variance, with π_k fixed to 1 in this work.The authors identify more optimal mixture parameters as future work.
- MeanFlow sampling: A lightweight decoder estimates the average velocity mapping from GMN samples to expert trajectories using scene context and encoded time intervals.During training, the nearest Gaussian component to the ground truth supplies the loss; during inference, components generate trajectories in parallel.
- Adaptive Reconstruction Module: ARM applies cross-attention to sampled candidates and scene features, then a Projector outputs the final trajectory under expert-trajectory supervision.Its reconstruction loss is combined with MeanFlow and mapping losses, and training does not use benchmark evaluation metrics.
5. Experiments
MeanFuser is evaluated on NAVSIM using multimodal trajectory generation, adaptive reconstruction, and one-step inference. It achieves strong closed-loop performance, high inference speed, diverse trajectory proposals, and improved robustness when proposals are suboptimal.
- Main Results: MeanFuser reaches PDMS 89.0 on NAVSIMv1 and EPDMS 89.5 on NAVSIMv2 under a shared ResNet-34 visual backbone using RGB input only.
- Ablation Studies: GMN and ARM add +0.9 PDMS and +0.8 PDMS, respectively, producing a final performance of 89.0 PDMS.Replacing ARM with proposal averaging causes a 17.8 PDMS drop.
- Ablation Studies: ARM reduces cases where all proposals leave the drivable area from 58 to 48 by regenerating trajectories when sampled proposals are suboptimal.GMN also captures richer driving-strategy diversity than standard Gaussian noise sampling.
- Trajectory Diversity: Sampling from Gaussian components produces diverse driving styles, while the model represents distinct strategies such as lane keeping and left-lane changing.The visualizations show multimodal trajectory behavior across Gaussian components and inferred strategies.
6. Conclusion
The paper presents MeanFuser as an end-to-end planning framework designed to balance one-step sampling efficiency with trajectory-planning performance. Its core design combines continuous multimodal noise modeling, MeanFlow-based generation, and adaptive reconstruction.
- MeanFuser addresses discrete vocabulary dependence while balancing one-step sampling and planning performance.
- The framework combines Gaussian mixture noise, the MeanFlow identity, and an adaptive reconstruction module.The paper analyzes planning performance, computational efficiency, and trajectory diversity.
7. CARLA Longest6 Benchmark
MeanFuser is evaluated on the CARLA Longest6 closed-loop benchmark across routes, towns, weather conditions, and times of day. It outperforms both unimodal and multimodal comparison methods in Driving Score.
- Benchmark: CARLA Longest6 contains 36 test routes across six towns, averaging 1500 meters, with six weather conditions and six times of day.Its Driving Score combines Route Completion with an Infraction Score penalty.
- Results: MeanFuser surpasses TransFuser by 22.78 DS and DiffusionDrive by 5.81 DS in closed-loop testing.The reported results use three runs per route to compute mean and standard deviation.
8. Further Ablation Study
The ablations examine Gaussian Mixture Noise design choices, component counts, and multimodal trajectory quality. Results show that GMN configuration affects performance and diversity, with eight components providing the best reported performance and GMN improving the combined multimodality-performance metric.
- Gaussian component count: Eight Gaussian components achieve optimal performance, while adding more components does not improve results and may slightly reduce performance.The authors attribute this to sufficient trajectory-distribution capacity at eight components; further increases can leave components data-starved and produce unreliable velocity predictions.
- GMN generation: Manual GMN construction decreases performance by only 0.45% relative to clustering expert trajectories, whereas standard-normal components cause significant degradation.The comparison uses heuristic component means with a unified fixed standard deviation for manual construction.
- Multimodal planning: The multimodality metric D compares bounding-box intersection and union across K trajectories at each prediction timestep.Higher D indicates greater diversity among predicted trajectories, and the intersection-over-union dynamics are visualized across timesteps.
- Multimodal planning: GMN increases both primary planning performance and trajectory diversity, improving the combined MDP metric by 20.84%.The composite metric MDP combines the multimodality metric D with PDMS, while D measures diversity across predicted trajectories.