PaperlayerView exact revision on arXiv

Source-linked AI summary

GuideFlow: Constraint-Guided Flow Matching for Planning in End-to-End Autonomous Driving

Lin Liu, Caiyan Jia, Guanyi Yu, Ziying Song, JunQiao Li, Feiyang Jia, Peiliang Wu, Xiaoshuai Hao, Yadan Luo

arXiv:2511.18729v3cs.CV

TL;DR

Imitative planners can collapse to a dominant trajectory mode, while generative planners struggle to ensure safety constraints during generation. GuideFlow uses constrained flow matching with an energy-based model; it reports SOTA on Navhard with 43.0 EPDMS.

  • Problem

    Imitation-trained multimodal planners tend to collapse toward one dominant trajectory mode, while generative sampling makes it difficult to guarantee hard safety constraints.

    Source: §1 · p. 2

  • Method

    GuideFlow guides flow-matching trajectory generation with explicit constraints, conditioning signals, and an energy-based model.

    Source: §1 · p. 2, §4.1 · p. 4, §4.3 · p. 5

  • Results

    43.0 EPDMS on the NavSim test hard split (Navhard) is reported as SOTA, alongside evaluations on Bench2Drive, NuScenes, NavSim, and ADV-NuScenes.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The reported evaluations support GuideFlow’s effectiveness, with the conclusion noting robustness particularly in challenging out-of-domain scenarios.

    Source: §1 · p. 2, §7 · p. 9

Abstract

from arXiv · show

Driving planning is a critical component of end-to-end (E2E) autonomous driving. However, prevailing Imitative E2E Planners often suffer from multimodal trajectory mode collapse, failing to produce diverse trajectory proposals. Meanwhile, Generative E2E Planners struggle to incorporate crucial safety and physical constraints directly into the generative process, necessitating an additional optimization stage to refine their outputs. In this paper, we propose \textit{\textbf{GuideFlow}}, a novel planning framework that leverages Constrained Flow Matching. Concretely, \textit{\textbf{GuideFlow}} explicitly models the flow matching process, which inherently mitigates mode collapse and allows for flexible guidance from various conditioning signals. Our core contribution lies in directly enforcing explicit constraints within the flow matching generation process, rather than relying on implicit constraint encoding. Crucially, \textit{\textbf{GuideFlow}} unifies the training of the flow matching with the Energy-Based Model (EBM) to enhance the model's autonomous optimization capability to robustly satisfy physical constraints. Secondly, \textit{\textbf{GuideFlow}} parameterizes driving aggressiveness as a control signal during generation, enabling precise manipulation of trajectory style. Extensive evaluations on major driving benchmarks (Bench2Drive, NuScenes, NavSim and ADV-NuScenes) validate the effectiveness of \textit{\textbf{GuideFlow}}. Notably, on the NavSim test hard split (Navhard), \textit{\textbf{GuideFlow}} achieved SOTA with an EPDMS score of 43.0. The code will be in https://github.com/liulin815/GuideFlow.

1. Introduction

Imitative planners can collapse to a dominant trajectory mode, while generative planners have difficulty ensuring safety constraints during sampling. GuideFlow guides flow-matching generation with explicit constraints and conditioning signals.

1. Introduction

  • Imitation-trained multimodal planners tend to collapse toward one dominant mode because each scene supplies only one ground-truth trajectory, while generative sampling raises hard-constraint concerns.The introduction identifies explicit style and safety guidance during generation as rarely explored.Source: §1 · p. 2
  • GuideFlow uses flow matching with random initial samples and diverse conditioning signals, and explicitly supervises the generation process.The paper presents this as its approach to mitigating mode collapse.Source: §1 · p. 2
  • CVF corrects predicted velocities, CF corrects deviating flow paths, and RFE unifies flow matching with an EBM to explore constraint-satisfying results.Together, these strategies embed safety constraints in generation rather than relying only on implicit encoding.Source: §1 · p. 2
  • 43.0 EPDMS on the NavSim test hard split (Navhard) gives GuideFlow SOTA, and evaluations cover Bench2Drive, NuScenes, NavSim, and ADV-NuScenes.The paper reports this as its headline result.Source: §1 · p. 2
  • Environmental rewards condition generation to switch between aggressive and conservative driving styles during inference.Source: §1 · p. 2

2. Related Work

Prior imitative planners learn from a single demonstrated trajectory, whereas generative planners model distributions over future trajectories. GuideFlow combines explicit generative-process supervision with an energy-based model and direct hard-constraint injection.

2. Related Work

  • Imitative planners use single-trajectory supervision and can collapse to the dominant expert mode, limiting representation of multiple plausible intentions.The cited examples differ architecturally but use Huber/L1 imitation losses.Source: §2 · p. 2
  • Generative planners model trajectory distributions, but their sampling can remain constraint-unaware; GuideFlow instead combines process supervision, an EBM, and direct hard-constraint injection.The passage contrasts this with inference-time energy biasing and denoising methods whose sampling is not constraint-aware.Source: §2 · p. 2

3. Preliminary

Flow-based planning transports samples from a Gaussian prior toward a target trajectory distribution using a learned vector field. Rectified flow provides a straight transport path, while energy-based dynamics are described as supporting multiple feasible modes.

3. Preliminary

  • A learned vector field transports a Gaussian prior distribution toward the target trajectory distribution along a probability path.Source: §3 · p. 3
  • Rectified flow uses a linear path between prior and target samples and learns straight transport toward the data manifold.At inference, trajectories are generated by numerical integration.Source: §3 · p. 3
  • Straight transport is described as mode-seeking and prone to collapsing toward the dominant driving pattern; Energy Matching adds an energy function to recover multiple feasible modes.Source: §3 · p. 3
  • The described energy schedule transitions from flow transport to energy-guided manifold refinement, with energy shaping multiple low-energy basins such as yield and merge modes.The passage says the energy term activates after the transport phase, guiding samples into distinct low-energy modes.Source: §3 · p. 3

4. Methodology

GuideFlow generates trajectories through a perception-conditioned flow, adding intent and style signals and enforcing constraints during generation. Its constrained sampling combines velocity correction, flow-state replacement, and energy-based refinement.

4. Methodology

  • GuideFlow combines a perception-conditioned velocity field, intent and style guidance, and safety-constrained sampling to generate feasible and safe future motion plans.Its constrained sampling uses velocity-field correction, flow-state constraints, and energy-based refinement.Source: §4 · p. 3

4.1. Perception-conditioned Flow Generator

  • Scene tokens encode dynamic-agent interactions and road and lane topology, conditioning the flow velocity field that generates future trajectories.The perception module derives these tokens from BEV features extracted from multi-view images.Source: §4.1 · p. 4

4.2. Classifier-free Intent and Reward Guidance

  • GuideFlow conditions generation on a plan anchor, goal point, driving command, or reward, then applies classifier-free guidance to control their influence.The three intent signals overlap semantically and are not used simultaneously; plan anchors support diverse candidate motions.Source: §4.2 · p. 4

4.3. Constrained Generation

  • Velocity-field constraints steer the predicted direction toward a feasible reference while minimally changing velocity magnitude.The correction uses λ = 0.1.Source: §4.3 · p. 4
  • Flow-state constraints replace a discrete state near the target with a constraint-satisfying anchor and continue sampling, activating this truncation-like correction only at inference.The replacement step is set to k_c = 50 in practice.Source: §4.3 · p. 5
  • Energy-based refinement assigns lower energy to feasible trajectories and higher energy to constraint violations, training the velocity field toward constraint-satisfying regions.The energy objective raises energy for generated endpoints that violate constraints and lowers it for those that satisfy them.Source: §4.3 · p. 5

4.4. Reward as Style Condition

  • An online aggressiveness score based on lane-centerline distance traveled per unit time conditions generation, and setting it near 1 produces more aggressive behavior.The score ranges from 0 to 1.Source: §4.4 · p. 5

5. Experiments

GuideFlow is evaluated on open- and closed-loop driving benchmarks, with reported gains in planning scores and collision rates. Ablations examine conditioning signals, constraint modules, reward-based style control, and sampling hyperparameters.

5.1. Experimental Setup

  • GuideFlow uses image inputs for open-loop NuScenes and ADV-NuScenes evaluation, and closed-loop NavSim and Bench2Drive testing with benchmark-specific metrics.ADV-NuScenes is used only for out-of-domain evaluation and excluded from training; the NavSim and Bench2Drive setups align training protocols and baselines for comparison.Source: §5.1 · p. 5

5.2. Main Results

  • 43.0 EPDMS on Navhard with the Scorer sets a new SOTA, while GuideFlow reports 75.04 Driving Score and 50.90% Success Rate on Bench2Drive.The Navhard result exceeds the prior best by +1.3; GuideFlow also reaches 27.1 EPDMS there without the Scorer.Source: §5.2 · p. 6, Table 1 · p. 6, Table 2 · p. 6
  • 0.07% average collision rate on NuScenes and 0.73% on ADV-NuScenes are the lowest reported rates across prediction horizons.These results outperform SparseDrive by 0.08% and 1.02%, respectively; the passage also reports nearly zero collisions at 1s and 0.02% at 2s.Source: §5.2 · p. 6, Table 3 · p. 7

5.3. Ablation Study

  • In the reported dynamic-condition ablation, plan-anchor guidance achieves the highest performance among tested signals, supporting the efficacy of Classifier-free Intent and Reward Guidance relative to the baseline.Source: §5.3 · p. 6
  • +1.6 EPDMS and +0.45% Success Rate are reported for CF versus CVF, while CF and RFE together reach 27.1 EPDMS, 75.21 Driving Score, and 51.36% Success Rate.The passage says CF intervenes once, whereas CVF corrects at every generation step; RFE contributes most to EPDMS, particularly OOD Stage2 scoring, and the two modules are described as complementary.Source: §5.3 · p. 7, Table 5 · p. 8
  • Setting the EP reward to 1 raises EP from 79.6 to 82.3 with RAS, alongside a 0.8-point decrease in EPDMS.The ablation describes this as evidence that reward conditioning can modulate aggressiveness, while indiscriminate encouragement of aggressive trajectories compromises safety constraints.Source: §5.3 · p. 7, Table 5 · p. 8
  • Increasing λ from 0.1 to 0.5 decreases EPMDS, and increasing kc from 10 to 50 first raises then lowers EPMDS.The passage attributes the λ trend to excessive interference with the predicted velocity field; it says initiating constraints too late leaves insufficient steps for adaptation.Source: §5.3 · pp. 7–8, Table 6 · p. 8
  • Excessively large K disrupts sampling stability and can produce erratic trajectories and degraded performance despite rectified flow’s theoretically straight trajectories.The passage attributes this practical limit to deviations from the ideal model.Source: §5.3 · p. 8, Table 6 · p. 8

6. Qualitative Results

The visual comparison presents GuideFlow trajectories as more lane-following, smoother through maneuvers, and more compliant with collision and road-boundary constraints than DiffusionDrive.

6. Qualitative Results

  • GuideFlow exhibits improved lane-following and smoother maneuver transitions than DiffusionDrive across the pictured scenarios.The examples include lane-change and turn scenarios.Source: §6 · p. 8
  • In scenarios with surrounding vehicles, GuideFlow shows collision-avoidance maneuvers and stays stationary to avoid a potential collision with the leading vehicle.Source: §6 · p. 8
  • The comparison highlights stronger compliance with safety constraints, including collision avoidance and road-boundary preservation.Source: §6 · p. 8

7. Conclusion

GuideFlow uses flow matching for planning, combining diverse conditional signals with explicit constraint-enforcement strategies. Evaluations report effectiveness and robustness in challenging out-of-domain scenarios, while accelerated sampling can compromise performance.

7. Conclusion

  • GuideFlow leverages flow matching for planning and incorporates diverse conditional signals to guide context-aware trajectory generation.Signals include driving commands, goal points, and planning anchors.Source: §7 · pp. 8–9
  • GuideFlow proposes three distinct strategies to enforce explicit constraints throughout the generation process.Source: §7 · p. 9
  • Experiments across NavSim, NuScenes, and Bench2Drive confirm GuideFlow’s effectiveness and superior robustness in challenging out-of-domain scenarios.Source: §7 · p. 9
  • Accelerated sampling can compromise performance, and future work will integrate reflow and meanflow to improve sampling speed.Source: §7 · p. 9

Related papers

Loading 2511.18729v3…