FlowDrive: Autonomous Flow Planning

Updated 4 July 2026

FlowDrive is a family of autonomous driving planning systems that use flow-based generative modeling to predict future trajectories in either BEV space or trajectory space.
The energy-field variant utilizes explicit risk potential and lane attraction fields with anchor refinement and conditional diffusion to generate interpretable trajectories.
The rectified-flow variant employs data balancing and in-the-loop guidance by transforming Gaussian noise into diverse, reactive trajectory predictions.

FlowDrive is the name used by two 2025 autonomous-driving planning frameworks that apply flow-based generative modeling to trajectory planning, but they differ substantially in formulation and emphasis. One version, "FlowDrive: Energy Flow Field for End-to-End Autonomous Driving," introduces physically interpretable energy-based flow fields in bird’s-eye-view (BEV) space, including a risk potential field and a lane attraction field, together with a motion-decoupled conditional diffusion planner (Jiang et al., 17 Sep 2025). A second version, "FlowDrive: moderated flow matching with data balancing for trajectory planning," formulates planning as conditional rectified flow from Gaussian noise to future ego trajectories, and combines this with trajectory-pattern-based data balancing and moderated in-the-loop guidance (Wang et al., 26 Sep 2025). In current usage, the term therefore denotes an emerging family of flow-based planning systems for autonomous driving rather than a single canonical method.

1. Naming and scope

The shared name "FlowDrive" refers to two distinct planning architectures introduced in 2025 for end-to-end or learning-based autonomous driving (Jiang et al., 17 Sep 2025). Both works operate in the trajectory-planning setting, where a model predicts a future ego trajectory conditioned on scene context, but they instantiate different notions of "flow."

In the energy-field formulation, "flow" denotes dense spatial guidance in BEV induced by scalar energy landscapes. The method introduces a risk potential field and a lane attraction field, uses these fields to refine anchor trajectories, and couples them to a conditional diffusion planner with feature-level motion decoupling (Jiang et al., 17 Sep 2025). In the rectified-flow formulation, "flow" denotes a learned conditional transport map from a simple prior to the data distribution of future trajectories. The planner learns a conditional velocity field under a rectified-flow objective, then generates trajectories in a small number of Euler steps, with additional inference-time moderated guidance for diversity (Wang et al., 26 Sep 2025).

This naming overlap suggests that, within autonomous-driving research, "FlowDrive" has become associated with planners that make trajectory generation more structured, either by explicit field priors in BEV or by explicit transport dynamics in trajectory space. A plausible implication is that the term is better interpreted as a label for flow-centric planning design than as a uniquely identified architecture.

2. FlowDrive as energy flow fields in BEV

"FlowDrive: Energy Flow Field for End-to-End Autonomous Driving" addresses end-to-end planning in BEV by arguing that autonomous vehicles must reason about both hard geometric constraints, such as occupied regions corresponding to vehicles and pedestrians, and soft semantic or rule-based constraints, such as lane boundaries and lane-center priors (Jiang et al., 17 Sep 2025). The method therefore introduces explicit, physically interpretable energy flow fields rather than relying only on implicitly learned BEV features.

The core representation is a pair of dense scalar fields over BEV space. The first is a risk potential field,

$U_{\text{risk}(u,v) = \sum_i \eta \exp\left( -\frac{\|(u,v)-(u_i,v_i)\|^2}{2\sigma^2} \right),$

where $(u_i,v_i)$ denotes the BEV position of agent $i$ , $\eta$ is a risk weight, and $\sigma$ controls spatial spread (Jiang et al., 17 Sep 2025). The second is a lane attraction field, intended to penalize lateral deviation from the lane centerline and encode longitudinal progress along the lane. The paper’s printed equation is malformed, but its intended form uses the nearest-centerline distance $d(u,v)$ , longitudinal arc length $s(u,v)$ , total center-lane length $L$ , and weights $k_{\text{lat}}$ and $k_{\text{lon}}$ (Jiang et al., 17 Sep 2025). The paper repeatedly interprets the gradients of these fields as directions of energy descent, although it does not state an explicit formula such as $(u_i,v_i)$ 0.

Architecturally, the method begins with BEV feature construction using a TransFuser-style perception backbone, with auxiliary supervision from BEV semantic map segmentation and object detection (Jiang et al., 17 Sep 2025). Flow features are then learned from BEV features using learnable queries, multi-head attention, and feed-forward layers, producing $(u_i,v_i)$ 1, from which the model predicts

$(u_i,v_i)$ 2

These fields are supervised against analytically constructed target fields derived from scene annotations and lane priors (Jiang et al., 17 Sep 2025).

A second key component is flow-aware anchor refinement. The planner starts from $(u_i,v_i)$ 3 trajectory anchors obtained by K-means clustering over training trajectories, and predicts residual corrections conditioned on flow features and ego features:

$(u_i,v_i)$ 4

The printed notation is imperfect, but the intended mechanism is a learned scene-dependent update that aligns anchors with safer and more lane-consistent regions before generative planning (Jiang et al., 17 Sep 2025).

The final planner is a conditional diffusion model. The forward process is

$(u_i,v_i)$ 5

with $(u_i,v_i)$ 6, and inference uses a DDIM-style deterministic reverse update (Jiang et al., 17 Sep 2025). A distinctive design choice is feature-level motion decoupling via two sets of learnable gating queries $(u_i,v_i)$ 7 and $(u_i,v_i)$ 8, intended to separate motion intent prediction from trajectory denoising (Jiang et al., 17 Sep 2025).

This formulation makes FlowDrive notable as an attempt to hybridize learned BEV planning with explicit potential-field-like priors. The paper’s interpretability claim rests on the fact that obstacle repulsion and lane attraction are represented as inspectable dense fields rather than remaining implicit in latent tensors.

3. FlowDrive as moderated flow matching with data balancing

"FlowDrive: moderated flow matching with data balancing for trajectory planning" formulates planning as conditional generative modeling of future ego trajectories under a rectified-flow objective (Wang et al., 26 Sep 2025). Its starting point is the observation that learning-based planners are highly sensitive to the long-tailed structure of driving data. The paper reports, for roughly $(u_i,v_i)$ 9 sampled nuPlan training scenarios, that one may observe only about $i$ 0 sample of changing_lane_with_lead versus $i$ 1 stationary samples (Wang et al., 26 Sep 2025).

The model predicts future trajectories $i$ 2, with $i$ 3 corresponding to $i$ 4, conditioned on scene context $i$ 5 comprising HD map information, traffic lights, static objects, dynamic neighboring agents, route information, and current ego state (Wang et al., 26 Sep 2025). The paper reports $i$ 6 predicted poses. The planner models the conditional distribution $i$ 7 using a rectified-flow path

$i$ 8

with $i$ 9 and $\eta$ 0 (Wang et al., 26 Sep 2025). The target instantaneous velocity along the pair $\eta$ 1 is

$\eta$ 2

and the model learns $\eta$ 3 by minimizing

$\eta$ 4

At inference, trajectories are generated by integrating

$\eta$ 5

using Euler discretization (Wang et al., 26 Sep 2025).

The paper’s main training-time contribution is data balancing by trajectory pattern. It compares no weighted sampling, scenario-based sampling, and cluster-based sampling, where future ego trajectories are clustered with K-means using $\eta$ 6 on stacked future $\eta$ 7 positions, and inverse-frequency weights are assigned analogously to scenario-based reweighting (Wang et al., 26 Sep 2025). The reported ablation on nuPlan Val14 reactive is 81.91 with no weighted sampling, 80.08 with scenario-based sampling, and 85.37 with cluster-based sampling (Wang et al., 26 Sep 2025). The paper therefore argues that balancing by the distribution the planner actually predicts—future ego motion—is more effective than balancing by coarse scenario labels.

Its main inference-time contribution is moderated in-the-loop guidance. Rather than shifting trajectories only after generation, the method injects small perturbations into the positional component of the trajectory state during flow integration:

$\eta$ 8

where $\eta$ 9, $\sigma$ 0, $\sigma$ 1, and $\sigma$ 2 (Wang et al., 26 Sep 2025). Only position is modified; $\sigma$ 3 remains unchanged. The perturbation is inserted before the next velocity-field evaluation so that later flow steps can reconcile it with the scene (Wang et al., 26 Sep 2025).

The paper finds that a single injection step is sufficient, compares early, middle, and late injection at $\sigma$ 4, and selects mid-step injection at $\sigma$ 5 as the best trade-off (Wang et al., 26 Sep 2025). Longitudinal guidance does not improve closed-loop performance, so all experiments set $\sigma$ 6; lateral offsets $\sigma$ 7 are used to induce candidate diversity such as overtaking or nudging behavior (Wang et al., 26 Sep 2025).

This version of FlowDrive thus treats "flow" in the literal rectified-flow sense: a learned conditional transport from Gaussian noise to trajectory distributions. Its central practical claim is that few-step flow matching, when combined with trajectory-pattern balancing and controlled in-the-loop perturbation, yields a strong learning-based planner.

4. Comparative structure of the two FlowDrive formulations

The two FlowDrive systems share a common application domain—autonomous-driving trajectory planning—but differ in representation, inductive bias, and generative mechanism.

Formulation	Primary planning prior	Generative mechanism
Energy flow fields	Risk potential field and lane attraction field in BEV	Conditional diffusion planner with anchor refinement
Moderated flow matching	Rectified transport from Gaussian noise to trajectories	Conditional rectified flow with Euler integration

In the energy-field formulation, the dominant inductive structure is spatial and BEV-centric. Scene understanding is organized around dense energy landscapes over BEV, and planning is scaffolded by explicit risk and lane priors, then refined through anchor trajectories and a conditional diffusion process (Jiang et al., 17 Sep 2025). In the rectified-flow formulation, the dominant inductive structure is distributional and trajectory-centric. The planner directly models $\sigma$ 8 by transporting noise samples in trajectory space, while dataset imbalance is handled through weighted sampling and candidate diversity is induced during integration (Wang et al., 26 Sep 2025).

Their notions of interpretability also differ. The energy-field system is interpretable primarily through visualizable BEV fields: obstacle-induced repulsion and lane-induced attraction can be inspected directly (Jiang et al., 17 Sep 2025). The rectified-flow system is interpretable primarily through its training distribution and candidate-generation mechanism: trajectory clusters, flow steps, and moderated guidance expose how the model’s behavior depends on data balancing and injected lateral offsets (Wang et al., 26 Sep 2025).

A plausible implication is that the two works instantiate complementary design philosophies. One externalizes semantic priors as explicit BEV fields; the other externalizes generative dynamics and diversity control in trajectory space. The shared name masks a substantive methodological divergence.

5. Benchmarks and reported performance

The two FlowDrive papers evaluate on different benchmarks and report different headline metrics. The energy-field formulation is evaluated on NAVSIM v2 using EPDMS and related component metrics (Jiang et al., 17 Sep 2025). The rectified-flow formulation is evaluated on nuPlan and interPlan using benchmark driving scores in non-reactive and reactive simulator settings (Wang et al., 26 Sep 2025).

The main reported results can be summarized briefly.

Method	Benchmark	Headline result
Energy flow fields	NAVSIM v2	EPDMS $\sigma$ 9
Moderated flow matching	nuPlan Val14 reactive	$d(u,v)$ 0
FlowDrive*	Test14 reactive / interPlan	$d(u,v)$ 1 / $d(u,v)$ 2

For the energy-field formulation, the paper reports that FlowDrive achieves state-of-the-art EPDMS 86.3 on NAVSIM v2 with the V2-99 backbone, compared with 86.0 for DriveSuprim, 85.1 for HydraMDP++, and 85.0 for DiffusionDrive (Jiang et al., 17 Sep 2025). In the same setting, FlowDrive reports NC 98.5, DAC 97.4, DDC 99.6, TL 99.9, EP 87.7, TTC 97.9, LK 97.8, HC 98.3, and EC 87.6 (Jiang et al., 17 Sep 2025). An ablation shows that removing flow field learning reduces EPDMS from 86.3 to 85.8, removing anchor refinement reduces it to 85.9, and removing motion decoupling reduces it to 86.1, suggesting the largest contribution comes from explicit flow field learning (Jiang et al., 17 Sep 2025).

For the rectified-flow formulation, the paper reports on Val14 reactive that PlanTF scores 76.95, Diffusion Planner 82.38, FlowDrive $d(u,v)$ 3 81.91, and FlowDrive 85.37 (Wang et al., 26 Sep 2025). On Test14 reactive, FlowDrive reports 87.28; on interPlan, it reports 36.96 (Wang et al., 26 Sep 2025). The hybrid variant FlowDrive*, which adds moderated guidance and light post-processing, reaches 92.96 on Val14 reactive, 93.96 on Test14 reactive, and 44.05 on interPlan (Wang et al., 26 Sep 2025). The candidate-count ablation for FlowDrive* finds that 30 candidates perform best, with reported scores 0.9226, 0.9292, 0.9296, and 0.9258 for 10, 20, 30, and 40 candidates respectively (Wang et al., 26 Sep 2025).

The rectified-flow paper also reports efficiency figures. The best setting uses 2000 training flow steps and 8 inference flow steps; runtime is 40 ms per inference pass on a single NVIDIA GeForce RTX 2080 Ti GPU, and generating 30 guided trajectories in batch increases runtime only to 43 ms (Wang et al., 26 Sep 2025). By contrast, the energy-field paper emphasizes system design and benchmark performance rather than a direct runtime comparison, but reports training on 8 NVIDIA H20 GPUs for 100 epochs with batch size 8 and initial learning rate $d(u,v)$ 4 (Jiang et al., 17 Sep 2025).

Taken together, these results suggest that both FlowDrive variants are benchmark-oriented high-performance planners, but they optimize for different evaluation ecosystems: NAVSIM v2 in one case and nuPlan/interPlan in the other.

6. Relation to adjacent flow-based driving research and limitations

The two FlowDrive works sit within a broader 2025–2026 trend of flow-based autonomous-driving planning and world modeling. "DynFlowDrive: Flow-Based Dynamic World Modeling for Autonomous Driving" treats trajectory-conditioned latent scene transition as a rectified-flow problem and uses a stability-aware multi-mode trajectory selection strategy during training, while explicitly stating that the world model adds no inference overhead because it is not used at test time (Liu et al., 20 Mar 2026). "DriveAnchor: Progressive Anchor-based Flow Learning for Autonomous Driving Planning" also uses flow matching, but replaces the Gaussian prior with a vocabulary of 2,398 trajectory anchors, adds an Energy Field module for controllable anchor relocation, and applies zeroth-order reinforcement learning for safety refinement (Yan et al., 30 May 2026). These related systems indicate that flow-based driving methods are diverging along multiple axes: BEV energy fields, latent world dynamics, anchor-based flow decoding, and rectified-flow trajectory generation.

The limitations of the two FlowDrive variants also differ. The energy-field paper does not formalize a vector-field equation for the BEV flows, does not provide an explicit gating equation for motion decoupling, and does not specify a standard DDPM noise-prediction objective (Jiang et al., 17 Sep 2025). Its formulation also depends on map and lane priors, and its evaluation is on simulation benchmark data rather than reported real-world deployment (Jiang et al., 17 Sep 2025). The rectified-flow paper notes that raw FlowDrive trajectories can be jerky, that human demonstrations may violate speed limits, and that best benchmark performance requires smoothing, speed-limit enforcement, and PDM-based scoring in the FlowDrive* variant (Wang et al., 26 Sep 2025). It also reports that longer training can improve non-reactive score while eventually hurting reactive performance, indicating overfitting (Wang et al., 26 Sep 2025).

A common misconception would be to treat "FlowDrive" as a single settled architecture. The literature instead contains at least two materially different systems under that name, both centered on flow-based planning but with different technical commitments. Another misconception would be to assume that flow-based planning here always refers to the same mathematical object. In one case it denotes explicit BEV energy landscapes; in the other it denotes rectified transport dynamics in trajectory space.

This ambiguity suggests that future references to FlowDrive require paper-level disambiguation. For researchers, the practical distinction is straightforward: if the emphasis is interpretable BEV priors, anchor refinement, and diffusion planning, the relevant method is the energy-field FlowDrive (Jiang et al., 17 Sep 2025); if the emphasis is rectified flow, long-tail data balancing, and in-the-loop guidance, the relevant method is the trajectory-planning FlowDrive (Wang et al., 26 Sep 2025).