FlowS: One-Step Motion Prediction via Local Transport Conditioning
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.
Rosa: We're looking at the paper: "FlowS: One-Step Motion Prediction via Local Transport Conditioning".
Dev: Here's the gist of the paper. Generative motion prediction must satisfy three simultaneous requirements for real-world autonomy: high accuracy, diverse multimodal futures, and strictly bounded latency The single most informative sentence about the paper's subject and finding is that local transport conditioning makes one-step generative motion prediction practical for safety-critical autonomy.
Rosa: The core problem addressed Motion prediction for autonomous driving must satisfy three simultaneous requirements: high accuracy, multimodal coverage, and strictly bounded latency The essential challenge is clear: given the recent scene with spatial positions and movement dynamics, predict where all agents will be within the next few seconds.
Dev: Diffusion models meet the first two but violate the third, requiring tens to hundreds of denoising steps.
Rosa: This observation reveals a strategy that we believe is fundamental to few-step generative prediction: Single-step integration is accurate when the transport problem is conditioned to be local.
Title and authors: Rosa: Now we look at the title again: FlowS: One-Step Motion Prediction via Local Transport Conditioning. It really tells you exactly what the paper is trying to solve.
Dev: It’s about taking motion prediction, which usually requires a lot of steps, and making it one step by conditioning it on local transport problems so we can achieve high accuracy without the latency penalty.
Taro: I wonder if they are really conditioning it on *local* transport in a general sense, or just local to the agent’s immediate vicinity. That’s where I want to probe.
Rosa: They explain that when you condition the base distribution to lie near plausible futures, you reduce the problem from discovering a global mode and traversing a long displacement to just short-range refinement.
Dev: That sounds like they are using the prior as a way to prune the search space dramatically before the main flow model even has to work hard.
Taro: Pruning is good, but I need to know if that pruning isn't just hiding errors in low-probability scenarios. What happens when the world misbehaves outside those anchors?
Rosa: They address that by using a scene-conditioned learned prior that emits K calibrated anchor trajectories per agent, each already near a plausible future, which converts mode discovery into local correction.
Dev: So instead of sampling from a huge Gaussian noise distribution and hoping for the right path, the AI starts with K options that are already somewhat correct? That should be much more efficient.
Taro: But if those K anchors don't cover the right maneuver, we’re back to square one, and I don't want to waste time on something that fails spectacularly in a real-world scenario.
Rosa: The paper shows that this prior reduces mean transport distance by thirty-six percent compared to starting from a standard Gaussian base sample Z ∼ N (zero I), which is about one point two three versus zero point seven eight, according to Figure five.
Dev: That thirty-six percent reduction in transport distance is significant because it directly tackles the long-range displacement issue that causes those massive discretization errors in a single step prediction.
Taro: So the prior helps by shortening the path, but what about the actual velocity estimation? The paper mentions that without this prior, the displacement field alone needs a mean velocity of one point seven two to bridge that full Gaussian-to-data gap.
Rosa: That's because without anchors, you have to learn everything from scratch—both where the agent is going and how fast it’s moving. The prior helps by eliminating the mode discovery burden while the displacement field eliminates discretization error on that short transport.
Dev: It’s a dual mechanism working together; one component handles distance reduction, and the other handles fidelity of movement estimation during that short path. That sounds like a very solid engineering approach for real-time performance.
Taro: I’m still focused on the training signal itself; how do they ensure that these two components actually cooperate during the learning phase without introducing instability?
Rosa: They use Conditional Flow Matching, or CFM, which learns a time-indexed flow model vphi by matching it to a known velocity target along a family of conditional probability paths. This is trained on an optimal transport straightline path between Z and Y.
Dev: So the training target is fixed as the empirical displacement from the base sample to the ground truth future, constant along each Optimal Transport path, which gives them something stable to match against.
Taro: That sounds like a good anchor for stability, unlike some prior self-consistency methods that can suffer from high-variance bootstrap signals when they are on curved diffusion paths.
Rosa: Exactly. The paper notes that anchoring the displacement field at learned priors along straightline paths yields a stable, low-variance training target, which helps stabilize the prior formulation compared to other ways of doing it.
Dev: So the key finding here is that this specific combination of conditioning and consistency creates a very stable signal for training the flow model. It’s less sensitive to noise during the learning process.
Taro: That stability is crucial for deployment; we need models that are robust when they encounter unexpected inputs, not just when they are in a perfectly straight-line training environment.
The paper's summary: Rosa: So, summarizing the core of FlowS: One-Step Motion Prediction via Local Transport Conditioning is that diffusion models struggle with latency because they need many steps to be accurate and diverse, but they found a way around this by conditioning the problem.
Dev: They identified that single-step integration is accurate if the underlying transport problem is local, meaning we only need to worry about short-range corrections instead of long-range discoveries.
Taro: The mechanism they use for this localization is twofold: first, a scene-conditioned learned prior that gives us K anchor trajectories already close to plausible futures.
Rosa: That prior converts the mode discovery problem into a local correction task, essentially telling the model where it’s probably going to be before it even starts predicting.
Dev: And then they add the second part: a step-consistent displacement field that ensures that even though we only take one update at inference time, that single step still respects the accuracy of multi-step dynamics during training.
Taro: So you have a starting point based on plausible futures, and then a mechanism to ensure the next move is physically consistent with what would happen over several steps.
Rosa: Precisely. The synergy between these components—the prior reducing transport distance and the field absorbing curvature—is what leads to state of the art performance on the Waymo Open Motion Dataset.
Dev: They achieved a soft mAP of zero point four eight zero four and an mAP of zero point four seven zero three, both with single-step inference running at about seventy-five frames per second, which is quite fast for real-time use cases.
Taro: That speed combined with those scores shows that this technique can actually be practical for safety critical autonomy when we need fast answers.
Rosa: It confirms the idea that conditioning on local transport makes one-step generative motion prediction practical for safety critical autonomy, which is the central finding of FlowS: One-Step Motion Prediction via Local Transport Conditioning.
The paper's improvements: Dev: Let's talk about what they actually improved in terms of methodology. They introduced a few specific changes to how they handle the flow matching process.
Rosa: The first improvement is using the learned prior instead of starting from a standard Gaussian base sample Z ∼ N (zero I) <ref:2604.26065#pg1>. As we saw in Figure three, this drastically reduces transport distance from one point two three down to zero point seven eight, which is a thirty-six percent reduction in mean transport distance.
Taro: Thirty-six percent is substantial; that means the AI doesn't have to travel as far across the state space before it starts making meaningful corrections, which directly lowers the error accumulation.
Dev: And then they implemented a step-consistent displacement field, spsi(Xt, t, d, c) which enforces semigroup self-consistency by ensuring that one two-dimensional update matches two composed d updates.
Taro: That’s the part that addresses discretization error in the flow dynamics; it makes sure the single step is actually a good approximation of what would happen over several steps.
Rosa: And they showed how these two components work together: the prior reduces transport distance by thirty-six percent, while the displacement field absorbs marginal curvature in vphi via semigroup consistency.
Dev: That dual mechanism is key; one component handles the distance reduction, and the other handles the cumulative effect of curvature in vphi during training. It stabilizes the bootstrap signal during training and ensures that when you take that single step at inference time, it realizes the full transport in one pass.
Taro: So they are not just improving one part; they’re fixing two independent sources of error: reducing the distance traveled and absorbing how quickly things change over time in a curved path.
Rosa: Exactly. They also showed that adding just the prior alone improves mAP by plus zero point zero zero eight two, and adding the step-consistent field alone yields an improvement of plus zero point zero zero five six on the validation set for component ablation studies, which is quite telling about their individual contributions.
Dev: It’s a solid finding because it quantifies exactly how much each piece contributes to the final result so you can see where the gains are coming from.
Taro: I'm still looking at those limitations they mentioned—the paper says the learned prior’s K equals six mixture may not capture all tail distribution maneuvers, especially for cyclist U-turns. That tells us we still have a gap in coverage that needs work before this is ready for mission critical use.
Rosa: That limitation is explicitly stated in their evaluation on the Waymo Open Motion Dataset; they say transfer to datasets with different map representations or shorter horizons would require encoder adaptation and prior recalibration.
Conclusion: Dev: So, wrapping up the FlowS: One-Step Motion Prediction via Local Transport Conditioning paper, the main implication is that we can achieve high accuracy and low latency for motion prediction by using local transport conditioning.
Rosa: They did this by creating a system where the learned prior anchors plausible futures, drastically reducing transport distance and then using a consistent displacement field to ensure single-step accuracy is maintained.
Taro: For me, it means that we have a way to build systems that can make fast decisions on the road without needing massive amounts of iterative computation for every decision.
Dev: It moves the problem from requiring hundreds of denoising steps down to one deterministic forward pass at about seventy-five frames per second, which is what we need for real-time systems.
Rosa: We're seeing strong performance on the Waymo Open Motion Dataset with mAP of zero point four seven zero three and soft mAP of zero point four eight zero four, showing that this method works effectively in practice when tested outside the lab environment as well.
Taro: I just hope they can expand beyond those initial evaluation sets to show that this robustness holds up when the world gets genuinely messy and unpredictable.
Dev: That's the next challenge; moving from a single-step prediction to something that handles more complex interaction scenarios with joint prediction, which is a natural next step for this research.
Rosa: So, FlowS: One-Step Motion Prediction via Local Transport Conditioning provides a concrete method for achieving one-step generative motion prediction practical for safety critical autonomy by grounding the displacement field at learned priors along straightline paths.
Leandro Di Bella, Bruno Cornelis, Adrian Munteanu
Vrije Universiteit Brussel · imec
cs.RO
Submitted: 2026-04-28
Updated: 2026-10-28
Comments: 8 pages
Journal ref: IEEE Robotics and Automation Letters, vol. 11, no. 12, pp. 13805-13812, Dec. 2026
License: http://creativecommons.org/licenses/by-nc-nd/4.0/
Importance score: 89/100
The gist: Generative motion prediction must satisfy three simultaneous requirements for real-world autonomy: high accuracy, diverse multimodal futures…
Key concepts
- Local Transport Conditioning
- This strategy makes motion prediction practical by ensuring the underlying transport problem is local. Instead of predicting a long, complex movement in one go, the model is guided to only perform short corrections near plausible future trajectories. This reduces discretization errors and allows a single step to be highly accurate.
- Scene-Conditioned Learned Prior
- This component replaces the standard random starting point with learned anchor trajectories already close to realistic futures. It converts the difficult task of discovering the correct overall movement mode into a simpler task of local correction, significantly reducing the transport distance required for prediction.
- Step-Consistent Displacement Field (spsi)
- This field enforces a mathematical consistency rule: two small sequential steps must approximate one larger step. This mechanism ensures that even though only one step is taken during inference, it inherits the accuracy of multi-step predictions, stabilizing training and reducing error caused by marginal curvature.
Terminology
Summary
Generative motion prediction must satisfy three simultaneous requirements for real-world autonomy: high accuracy, diverse multimodal futures, and strictly bounded latency The single most informative sentence about the paper's subject and finding is that local transport conditioning makes one-step generative motion prediction practical for safety-critical autonomy.
The core problem addressed
Motion prediction for autonomous driving must satisfy three simultaneous requirements: high accuracy, multimodal coverage, and strictly bounded latency The essential challenge is clear: given the recent scene with spatial positions and movement dynamics, predict where all agents will be within the next few seconds. Diffusion models meet the first two but violate the third, requiring tens to hundreds of denoising steps. p. 0 FlowS: One-Step Motion Prediction via Local
Transport Conditioning
Leandro Di Bella1
, Bruno Cornelis1
, and Adrian Munteanu1 Member, IEEE
Abstract—Generative motion prediction must satisfy three
simultaneous requirements for real-world autonomy: high accuracy, diverse multimodal futures, and strictly bounded latency. Diffusion models meet the first two but violate the third,
requiring tens to hundreds of denoising steps. Efforts to speed up sampling, such as DiffusionDrive [8] (truncated schedules from learned anchors) and
BridgeDrive [9] (diffusion bridges with matched PF-ODE
sampling), have reduced steps but still require multiple refinement cycles and cannot guarantee single-cycle latency. This
observation reveals a strategy that we believe is fundamental
to few-step generative prediction: Single-step integration is
accurate when the transport problem is conditioned to be
local.Where the paper says it
Where the paper says it
The conditioning strategy
Conditioning the base distribution to lie near plausible futures reduces the transport to a short local correction, the regime in which single-step integration is accurate. FlowS implements this with two complementary mechanisms:
-
Scene-conditioned learned prior. p. 0
Rather than starting from N (0, I), we learn an online, scene-conditioned multimodal prior that emits K calibrated anchor trajectories per agent, each already near a plausible future This converts the global mode-discovery problem into local correction. p. 0Where the paper says it
We instantiate this local transport conditioning in FlowS, a conditional flow matching framework with two mechanisms. First, an online, scene-conditioned learned prior emits K calibrated anchor trajectories per agent, each already near a plausible future, converting mode discovery into local correction.
Where the paper says it
Rather than starting from N (0, I), we learn an online, scene-conditioned multimodal prior that emits K calibrated anchor trajectories per agent, each already near a plausible future. This converts the global mode-discovery problem into local correction, satisfying the locality condition.
-
Step-consistent displacement field. p. 0
p. 0Where the paper says it
This converts the global mode-discovery problem into local correction, satisfying the locality condition. • Step-consistent displacement field.
We train a semigroup-consistent transport field that enforces T∆t ◦ T∆t ≈ T2∆t, so that a single step faithfully approximates multi-step dynamics. p. 0Where the paper says it
This converts the global mode-discovery problem into local correction, satisfying the locality condition. • Step-consistent displacement field.
Where the paper says it
• Step-consistent displacement field. We train a semigroup-consistent transport field that enforces T∆t ◦ T∆t ≈ T2∆t, so that a single step faithfully approximates multi-step dynamics.
The training signal is built
The training involves Conditional Flow Matching (CFM) which learns a time-indexed flow model vϕ by matching it to a known velocity target along a family of conditional probability paths. A convenient choice is the optimal transport (OT) straightline path between Z and Y. p. 2 CFM learns a time-indexed flow
model vϕ by matching it to a known velocity target along
a family of conditional probability paths {pt(· |Z, Y )}t∈[0,1]
that smoothly interpolate between p0 (at t=0) and pdata (at
t=1):
LCFM = EZ∼p0, Y ∼pdata
t∼U[0,1]
vϕ
Xt, t, c
− ut
Xt |Z, Y
2
2
. (1)
A convenient choice is the optimal transport (OT) straightline [3], [10] path between Z and Y :
Xt = (1 − t)Z + t Y, ut
Xt |Z, Y
= Y − Z, (2)
which yields a time-independent target velocity ut and
closed-form supervision. Substituting (2) into (1) gives:
LCFM = EZ,Y,t
vϕ
(1−t)Z +tY, t, c
−(Y −Z)
2
2
. (3)
In a one-step regime, the flow is applied once from the base
sample (t=0):
Ypred = Z + vϕ(Z, 0, c), (4)
C. A mixture calibration loss aligns predicted
weights π with soft targets πˆk = softmax(−NLLk/τmix) via
KL divergence:
Lmix = DKLπˆ ∥ π
=
XK
k=1
πˆk log(ˆπk/πk). (11)
The prior defines a K-component mixture distribution over
anchor trajectories; at inference, rather than sampling a single
component, we evaluate all K modes in parallel and use the
mixture weights πk as initial confidence scores for downstream
selection via NMS.Where the paper says it
Where the paper says it
Where the paper says it
Where the gains come from
The synergy between components is key: The prior reduces mean transport distance by 36% (0.78 vs. 1.23), while the displacement field alone must bridge the full Gaussian-to-data gap (mean velocity 1. p. 6 23), but a single Euler
step still accumulates discretization error without scale-aware
correction. Conversely, the displacement field alone must
bridge the full Gaussian-to-data gap (mean velocity 1. 72),
making its self-consistency target noisy. Together, the prior
eliminates the mode-discovery burden while the displacement
field eliminates discretization error on the resulting short
transport (mean velocity 1. latency (red, right
axis) across 1–16 inference steps. The semigroup constraint distills multi-step
accuracy into a single step: ∆mAP = 0. The semigroup constraint distills multi-step
accuracy into a single step: ∆mAP = 0. 0017 with 9. 0017 with 9. 5× speedup.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Performance and validation
On the Waymo Open Motion Dataset, FlowS (Ens.) achieves the highest Soft mAP (0. p. 5 BEST IN BOLD, SECOND UNDERLINED, THIRD MARKED ∗. FLOWS ACHIEVES THE HIGHEST MAP AND SOFT
MAP WITH SINGLE-STEP INFERENCE (∼75 FPS). On the Waymo
Open Motion Dataset, FlowS achieves state-of-the-art Soft mAP
(0. 4804) and mAP (0. On the Waymo
Open Motion Dataset, FlowS achieves state-of-the-art Soft mAP
(0. 4804) and mAP (0. 4804) and mAP (0. 4703), surpassing all
published entries including ensemble methods, while maintaining competitive displacement and miss rate. 4804) and mAP (0. 4703) with ensemble at 75 FPS with singlestep inference, demonstrating that local transport conditioning
makes one-step generative motion prediction practical for safetycritical autonomy. Without either, baseline Gaussian flow matching achieves 0. Adding the prior alone improves mAP by +0. 0082 (shorter
transport distance enables better mode capture). Adding the
step-consistent field alone yields +0. 25 ms/scene, a 34% reduction over MTR. The ensemble
aggregates predictions from 6 independently trained models
with varied hyperparameters (dropout rates, hidden dimensions, flow settings).Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Limitations and future work
The learned prior’s K = 6 mixture may not capture all taildistribution maneuvers (e.g., rare cyclist U-turns), especially true for cyclist maneuvers. p. 6 LIMITATIONS
The learned prior’s K = 6 mixture may not capture all taildistribution maneuvers (e. , rare cyclist U-turns), especially
true for cyclist maneuvers. While scaling K is a
good solution, it increases memory linearly. Our evaluation
is limited to WOMD; transfer to datasets with different map
representations, shorter horizons, or different agent densities
would require encoder adaptation and prior recalibration. The
current formulation predicts marginal per-agent futures without explicit joint consistency, which may produce overlapping
trajectories in dense interactions (though overlap rate remains
competitive; Tab. Extending FlowS to joint prediction via
interaction-conditioned flows is a natural next step. 58
displacement field that ensures one-step accuracy. Anchoring
the displacement field at learned priors along straight-line
paths yields a stable, low-variance self-consistency target, reducing the training instability of prior formulations on curved
diffusion paths.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
--- Page 1 ---
FlowS: One-Step Motion Prediction via Local Transport Conditioning Leandro Di Bella, Bruno Cornelis, and Adrian Munteanu Member, IEEE Abstract—Generative motion prediction must satisfy three simultaneous requirements for real-world autonomy: high accuracy, diverse multimodal futures, and strictly bounded latency Diffusion models meet the first two but violate the third, requiring tens to hundreds of denoising steps. We identify a conditioning strategy that resolves this tension: single-step integration is accurate when the underlying transport problem is local. p. 0 Diffusion models meet the first two but violate the third,
requiring tens to hundreds of denoising steps. We identify
a conditioning strategy that resolves this tension: single-step
integration is accurate when the underlying transport problem is
local. We identify
a conditioning strategy that resolves this tension: single-step
integration is accurate when the underlying transport problem is
local. A model that must both discover the correct behavioral
mode and traverse a long displacement in one step accumulates
large discretization errors; conditioning the base distribution to
lie near plausible futures reduces the problem to short-range
refinement, the regime where a single Euler step suffices. A model that must both discover the correct behavioral
mode and traverse a long displacement in one step accumulates
large discretization errors; conditioning the base distribution to
lie near plausible futures reduces the problem to short-range
refinement, the regime where a single Euler step suffices. We
instantiate this local transport conditioning in FlowS, a conditional
flow matching framework with two mechanisms. We
instantiate this local transport conditioning in FlowS, a conditional
flow matching framework with two mechanisms. First, an online, scene-conditioned learned prior emits K calibrated anchor
trajectories per agent, each already near a plausible future,
converting mode discovery into local correction. First, an online, scene-conditioned learned prior emits K calibrated anchor
trajectories per agent, each already near a plausible future,
converting mode discovery into local correction. Second, a stepconsistent displacement field enforces semigroup self-consistency,
guaranteeing that a single step inherits multi-step accuracy. Second, a stepconsistent displacement field enforces semigroup self-consistency,
guaranteeing that a single step inherits multi-step accuracy. Crucially, anchoring this field at learned priors along straightline paths yields a stable, low-variance training target, unlike
prior self-consistency methods that suffer from high-variance
bootstrap signals on curved diffusion paths. Crucially, anchoring this field at learned priors along straightline paths yields a stable, low-variance training target, unlike
prior self-consistency methods that suffer from high-variance
bootstrap signals on curved diffusion paths. On the Waymo
Open Motion Dataset, FlowS achieves state-of-the-art Soft mAP
(0. 4804) and mAP (0. 4703) with ensemble at 75 FPS with singlestep inference, demonstrating that local transport conditioning
makes one-step generative motion prediction practical for safetycritical autonomy.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
--- Page 2 ---
Fig. 1. FlowS architecture. p. 1 FlowS architecture. Scene inputs are processed by a Transformer-based encoder (Agent/Map). FlowS architecture. Scene inputs are processed by a Transformer-based encoder (Agent/Map). Scene inputs are processed by a Transformer-based encoder (Agent/Map). The Learned Prior module replaces the standard
Gaussian base by producing K scene-conditioned anchor trajectories. The Learned Prior module replaces the standard
Gaussian base by producing K scene-conditioned anchor trajectories. During training, the decoder (Dθ) learns flow dynamics from anchors to ground-truth
distributions via a step-consistent displacement loss (Lstep). During training, the decoder (Dθ) learns flow dynamics from anchors to ground-truth
distributions via a step-consistent displacement loss (Lstep). At inference, a single forward evaluation produces multi-agent trajectory predictions, filtered by
NMS.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
--- Page 3 ---
Fig. 2. Why a learned prior helps one-step flow matching. p. 2 Why a learned prior helps one-step flow matching. Left: Gaussian start. p. 2 Why a learned prior helps one-step flow matching. Left: Gaussian start. Left: Gaussian start. A scene-agnostic base sample Z ∼ N (0, I) can lie far from the target
future, so a single update must both identify the correct mode and traverse a long displacement, which increases error. A scene-agnostic base sample Z ∼ N (0, I) can lie far from the target
future, so a single update must both identify the correct mode and traverse a long displacement, which increases error. Right: learned prior. Right: learned prior. Scene-conditioned
anchors Ak already lie near plausible futures, so one-step CFM only needs a short corrective transport to reach the target. Scene-conditioned
anchors Ak already lie near plausible futures, so one-step CFM only needs a short corrective transport to reach the target. The prior therefore improves
conditioning by reducing transport length, yielding ∥YGT − Ak∥ ≪ ∥YGT − Z∥.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
--- Page 4 ---
Fig. 3. Step-consistent displacement field. Left: the base CFM velocity field vϕ(Xt, t, c) along the transport path from anchor A to the target. p. 3 Step-consistent displacement field. Left: the base CFM velocity field vϕ(Xt, t, c) along the transport path from anchor A to the target. Left: the base CFM velocity field vϕ(Xt, t, c) along the transport path from anchor A to the target. Middle:
semigroup consistency enforces that one 2d update matches two composed d updates. Middle:
semigroup consistency enforces that one 2d update matches two composed d updates. Right: at inference, the displacement field sψ(A, 0, 1, c) realizes the
full transport in one pass, while preserving the infinitesimal limit limd→0 sψ = vϕ.Where the paper says it
Where the paper says it
Where the paper says it
--- Page 5 ---
Fig. 4. Prior–displacement field synergy. p. 4 The displacement field sψ
addresses marginal curvature: as noted earlier, it absorbs the
cumulative effect of curvature in vϕ via semigroup consistency,
Fig. Prior–displacement field synergy. Prior–displacement field synergy. Left: Minimum transport distance
to ground truth for the learned prior (blue, mean 0. Left: Minimum transport distance
to ground truth for the learned prior (blue, mean 0. Gaussian N (0, I)
(red, mean 1. Gaussian N (0, I)
(red, mean 1. 23), showing a 36% reduction that enables accurate one-step
integration. 23), showing a 36% reduction that enables accurate one-step
integration. Right: One-step velocity magnitude for prior+displacement field
(blue, mean 1. Right: One-step velocity magnitude for prior+displacement field
(blue, mean 1. displacement field only from N (0, I) (red, mean 1.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
--- Page 6 ---
TABLE I WOMD TEST LEADERBOARD RESULTS BEST IN BOLD, SECOND UNDERLINED, THIRD MARKED ∗ FLOWS ACHIEVES THE HIGHEST MAP AND SOFT MAP WITH SINGLE-STEP INFERENCE (∼75 FPS) Method Venue Soft mAP ↑ mAP ↑ minADE ↓ minFDE ↓ Miss Rate ↓ Overlap Rate ↓ RMP-YOLO (Ens.) [25] ICRA’25 0. p. 5 Method Venue Soft mAP ↑ mAP ↑ minADE ↓ minFDE ↓ Miss Rate ↓ Overlap Rate ↓
RMP-YOLO (Ens. ) [25] ICRA’25 0. Method Venue Soft mAP ↑ mAP ↑ minADE ↓ minFDE ↓ Miss Rate ↓ Overlap Rate ↓
RMP-YOLO (Ens. ) [25] ICRA’25 0. ) [25] ICRA’25 0. 1259∗
ModeSeq (Ens. ) [25] ICRA’25 0. 1259∗
ModeSeq (Ens. 1259∗
ModeSeq (Ens. ) [26] CVPR’25 0. 1259∗
ModeSeq (Ens. ) [26] CVPR’25 0. ) [26] CVPR’25 0. 1275
BeTop [27] NeurIPS’24 0. 1275
BeTop [27] NeurIPS’24 0. 1272
MGTR [28] ICRA’24 0. 1259∗
ModeSeq (Ens. ) [26] CVPR’25 0. 1259∗
ModeSeq (Ens. ) [26] CVPR’25 0. 1272
MGTR [28] ICRA’24 0. 1275
EDA [29] AAAI’24 0. 1272
MGTR [28] ICRA’24 0. 1275
EDA [29] AAAI’24 0. 1272
MGTR [28] ICRA’24 0. 1275
EDA [29] AAAI’24 0. 1272
MGTR [28] ICRA’24 0. 1275
EDA [29] AAAI’24 0. 1272
MGTR [28] ICRA’24 0. 1275
EDA [29] AAAI’24 0. 1275
EDA [29] AAAI’24 0. 1266
MTR++ [15] TPAMI’24 0. 1275
EDA [29] AAAI’24 0. 1266
MTR++ [15] TPAMI’24 0. 1275
EDA [29] AAAI’24 0. 1266
MTR++ [15] TPAMI’24 0. 1275
EDA [29] AAAI’24 0. 1266
MTR++ [15] TPAMI’24 0. Method Venue Soft mAP ↑ mAP ↑ minADE ↓ minFDE ↓ Miss Rate ↓ Overlap Rate ↓
RMP-YOLO (Ens. ) [25] ICRA’25 0. Method Venue Soft mAP ↑ mAP ↑ minADE ↓ minFDE ↓ Miss Rate ↓ Overlap Rate ↓
RMP-YOLO (Ens. ) [25] ICRA’25 0. 1266
MTR++ [15] TPAMI’24 0. 1281
MTR [14] NeurIPS’22 0. ) [26] CVPR’25 0. 1275
BeTop [27] NeurIPS’24 0. ) [26] CVPR’25 0. 1275
BeTop [27] NeurIPS’24 0. ) [26] CVPR’25 0. 1275
BeTop [27] NeurIPS’24 0. 1281
MTR [14] NeurIPS’22 0. 1274
HPTR [30] AAAI’24 0. 1281
MTR [14] NeurIPS’22 0. 1274
HPTR [30] AAAI’24 0. 1272
MGTR [28] ICRA’24 0. 1275
EDA [29] AAAI’24 0. Method Venue Soft mAP ↑ mAP ↑ minADE ↓ minFDE ↓ Miss Rate ↓ Overlap Rate ↓
RMP-YOLO (Ens. ) [25] ICRA’25 0. Method Venue Soft mAP ↑ mAP ↑ minADE ↓ minFDE ↓ Miss Rate ↓ Overlap Rate ↓
RMP-YOLO (Ens. ) [25] ICRA’25 0. Method Venue Soft mAP ↑ mAP ↑ minADE ↓ minFDE ↓ Miss Rate ↓ Overlap Rate ↓
RMP-YOLO (Ens. ) [25] ICRA’25 0.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
--- Page 7 ---
TABLE II PER-CLASS PERFORMANCE ON THE VALIDATION SET OF WAYMO OPEN MOTION DATASET Setting Category mAP ↑ Vehicle 0.4756 Pedestrian 0. p. 0 The essential challenge
is clear: given the recent scene with spatial positions and
movement dynamics, predict where all agents will be within
the next few seconds. A forecast delivered after the next
control update is functionally useless: if a pedestrian steps
off the curb, the braking decision cannot wait. 6707
Pedestrian 0. 3551
Cyclist 0. 3551
Cyclist 0. 6969
Avg 0. Together, they achieve 0. 4508 mAP, the gains are
complementary, not redundant.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
--- Page 8 ---
TABLE III COMPONENT ABLATION (SINGLE-STEP INFERENCE, VALIDATION SET) Prior Step-Consistent Field mAP ↑ minADE ↓ × × 0.4376 0.5916 × 0.4432 0.5850 × 0.4458 0.5823 0.4508, Avg, Avg
--- Page 9 ---
TABLE IV DRIFT ANALYSIS BY SCENE DENSITY MINADE/MINFDE (M) ON THE WOMD VALIDATION SET Sparse (1–3) 0.943 0.983 1.208 1.265 Medium (4–5) 0. p. 6 Sparse (1–3) 0. 265 532
Medium (4–5) 0. 3551
Cyclist 0. 6969
Avg 0. 3551
Cyclist 0. 6969
Avg 0. 3551
Cyclist 0. 6969
Avg 0.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
--- Page 10 ---
TABLE V SCALE-AWARE DISPLACEMENT ANALYSIS DISPLACEMENT MAGNITUDE TARGET MAGNITUDE AND COSINE ALIGNMENT AT DIFFERENT POINTS ALONG THE FLOW PATH t d ∥sψ∥ ∥tgt∣cos(sψ, tgt) 0.00 1.00 0.729 0.730 16
--- Page 11 ---
TABLE VI PER-HORIZON ERROR ANALYSIS DISPLACEMENT ERROR (M) AT KEY PREDICTION HORIZONS ON THE WOMD VALIDATION SET Horizon Mean Median 95th Cum. ADE Horizon Mean Median 95th Cum. p. 7 DISPLACEMENT ERROR (M) AT KEY
PREDICTION HORIZONS ON THE WOMD VALIDATION SET. Horizon Mean Median 95th 99th Cum. DISPLACEMENT ERROR (M) AT KEY
PREDICTION HORIZONS ON THE WOMD VALIDATION SET. Horizon Mean Median 95th 99th Cum.Where the paper says it
Where the paper says it
--- Page 12 ---
REFERENCES [3] Y. Lipman, R. T. Chen, H. Ben-Hamu, M. p. 7 16948, 2023. Ben-Hamu, M. Ben-Hamu, M. Nickel, and M. Nickel, and M. Le, “Flow
matching for generative modeling,” arXiv preprint arXiv:2210. 02747,
2022. Cornman, C. Anguelov
et al. , “Motiondiffuser: Controllable multi-agent motion prediction using
diffusion,” in Proceedings of the IEEE/CVF conference on computer
vision and pattern recognition, 2023, pp. , “Motiondiffuser: Controllable multi-agent motion prediction using
diffusion,” in Proceedings of the IEEE/CVF conference on computer
vision and pattern recognition, 2023, pp. 9644–9653. Wolf, and Y. Bengio, “Conditional flow matching: Simulation-free
dynamic optimal transport,” arXiv preprint arXiv:2302. Wolf, and Y. Bengio, “Conditional flow matching: Simulation-free
dynamic optimal transport,” arXiv preprint arXiv:2302. Wolf, and Y. Bengio, “Conditional flow matching: Simulation-free
dynamic optimal transport,” arXiv preprint arXiv:2302. Wolf, and Y. Bengio, “Conditional flow matching: Simulation-free
dynamic optimal transport,” arXiv preprint arXiv:2302. 17 207–17 218. Chen, and Y. 17 207–17 218. Chen, and Y. 17 207–17 218. Chen, and Y. Zhang et al. , “Diffusiondrive: Truncated diffusion model
for end-to-end autonomous driving,” in Proceedings of the Computer
Vision and Pattern Recognition Conference, 2025, pp. Zhang et al. , “Diffusiondrive: Truncated diffusion model
for end-to-end autonomous driving,” in Proceedings of the Computer
Vision and Pattern Recognition Conference, 2025, pp. Zhang et al. , “Diffusiondrive: Truncated diffusion model
for end-to-end autonomous driving,” in Proceedings of the Computer
Vision and Pattern Recognition Conference, 2025, pp.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
--- Page 13 ---
TABLE VI PER-HORIZON ERROR ANALYSIS DISPLACEMENT ERROR (M) AT KEY PREDICTION HORIZONS ON THE WOMD VALIDATION SET Horizon Mean Median 95th Cum. ADE Horizon Mean Median 95th Cum. ADE
--- Page 14 ---
REFERENCES [3] Y. Lipman, R. T Chen et al. p. 7 17 207–17 218. Chen, and Y. 17 207–17 218. Chen, and Y. 17 207–17 218. Chen, and Y. 17 207–17 218. Chen, and Y. 17 207–17 218. Chen, and Y. Nickel, and M. Le, “Flow
matching for generative modeling,” arXiv preprint arXiv:2210. Nickel, and M. Le, “Flow
matching for generative modeling,” arXiv preprint arXiv:2210. Nickel, and M. Le, “Flow
matching for generative modeling,” arXiv preprint arXiv:2210. Nickel, and M. Le, “Flow
matching for generative modeling,” arXiv preprint arXiv:2210. Nickel, and M. Le, “Flow
matching for generative modeling,” arXiv preprint arXiv:2210. Le, “Flow
matching for generative modeling,” arXiv preprint arXiv:2210. 02747,
2022. Le, “Flow
matching for generative modeling,” arXiv preprint arXiv:2210. 02747,
2022. Le, “Flow
matching for generative modeling,” arXiv preprint arXiv:2210. 02747,
2022.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
--- Page 15 ---
TABLE V SCALE-AWARE DISPLACEMENT ANALYSIS DISPLACEMENT MAGNITUDE TARGET MAGNITUDE AND COSINE ALIGNMENT AT DIFFERENT POINTS ALONG THE FLOW PATH t d∥sψ∥∥tgt∣cos(sψ, tgt) 0.00 1.00 0.729 0.730
--- Page 16 ---
TABLE VI PER-HORIZON ERROR ANALYSIS DISPLACEMENT ERROR (M) AT KEY PREDICTION HORIZONS ON THE WOMD VALIDATION SET Horizon Mean Median 95th Cum. ADE
--- Page 17 ---
REFERENCES [3] Y. Lipman, R. T Chen et al., “Flow matching for generative modeling” arXiv preprint arXiv:2210.02747 [11] A Tong, N Malkin et al., “Conditional flow matching: Simulation-free dynamic optimal transport” arXiv preprint arXiv:2302.00482 [8] B Liao, S Chen et al., “Diffusiondrive: Truncated diffusion model for end-to-end autonomous driving” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2025, pp. 12 037–12 047
--- Page 18 ---
REFERENCES [3] Y. Lipman, R. T Chen et al., “Flow matching for generative modeling” arXiv preprint arXiv:2210.02747 [11] A Tong, N Malkin et al., “Conditional flow matching: Simulation-free dynamic optimal transport” arXiv preprint arXiv:2302.00482
--- Page 19 ---
REFERENCES [3] Y. Lipman, R. T Chen et al., “Flow matching for generative modeling” arXiv preprint arXiv:2210.02747
--- Page 26 ---
REFERENCES [3] Y. Lipman, R. T Chen et al., “Flow matching for generative modeling” arXiv preprint arXiv:2210.02747
--- Page 5 ---
Fig. 5. Quality–latency trade-off. p. 5 To assess training
Fig. Quality–latency trade-off. Quality–latency trade-off. mAP (blue, left axis) vs. mAP (blue, left axis) vs. latency (red, right
axis) across 1–16 inference steps. 5× speedup. Qualitative results on WOMD.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
--- Page 6 ---
TABLE I WOMD TEST LEADERBOARD RESULTS BEST IN BOLD, SECOND UNDERLINED, THIRD MARKED ∗ FLOWS ACHIEVES THE HIGHEST MAP AND SOFT MAP WITH SINGLE-STEP INFERENCE (∼75 FPS) Method Venue Soft mAP ↑ mAP ↑ minADE ↓ minFDE ↓ Miss Rate ↓ Overlap Rate ↓ RMP-YOLO (Ens.) [25] ICRA’25 0.4737 0.4531 0.5564∗ 1.1188 0.1084 0.1259∗ ModeSeq (Ens.) [26] CVPR’25 0.4737, CVPR’25) 0. p. 5 1259∗
ModeSeq (Ens. ) [26] CVPR’25 0. 1259∗
ModeSeq (Ens. ) [26] CVPR’25 0. Method Venue Soft mAP ↑ mAP ↑ minADE ↓ minFDE ↓ Miss Rate ↓ Overlap Rate ↓
RMP-YOLO (Ens. ) [25] ICRA’25 0. ) [26] CVPR’25 0. 1275
BeTop [27] NeurIPS’24 0. ) [26] CVPR’25 0. 1275
BeTop [27] NeurIPS’24 0. ) [26] CVPR’25 0. 1275
BeTop [27] NeurIPS’24 0. Method Venue Soft mAP ↑ mAP ↑ minADE ↓ minFDE ↓ Miss Rate ↓ Overlap Rate ↓
RMP-YOLO (Ens. ) [25] ICRA’25 0. 1275
BeTop [27] NeurIPS’24 0. 1272
MGTR [28] ICRA’24 0. 1259∗
ModeSeq (Ens. ) [26] CVPR’25 0. 1259∗
ModeSeq (Ens. ) [26] CVPR’25 0. 1259∗
ModeSeq (Ens. ) [26] CVPR’25 0. 1272
MGTR [28] ICRA’24 0. 1275
EDA [29] AAAI’24 0. 1275
EDA [29] AAAI’24 0. 1266
MTR++ [15] TPAMI’24 0. 1275
EDA [29] AAAI’24 0. 1266
MTR++ [15] TPAMI’24 0. 1266
MTR++ [15] TPAMI’24 0. 1281
MTR [14] NeurIPS’22 0. ) [26] CVPR’25 0. 1275
BeTop [27] NeurIPS’24 0.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Improved one-step generative motion prediction by conditioning it on local transport problems, which reduces discretization error when an agent must discover a correct behavioral mode and traverse a long displacement in one step. This allows for high accuracy at real-time speeds like 75 FPS using single-step inference.
-
Enhanced scene context representation by using a symmetric MTR++ encoder that vectorizes agent histories and HD-map polylines into polyline tokens processed with query centric self attention in local coordinates, preserving locality and scaling to large neighborhoods. This results in per agent context features c that condition all subsequent modules effectively.
-
Introduced a learned scene conditioned prior that emits K calibrated anchor trajectories per agent, which converts the
mode discovery problem into local correction
by ensuring anchors lie near plausible futures. This shortens transport distance, making the problem ashort-range refinement,
where a single Euler step suffices for accurate prediction. -
Implemented a step consistent displacement field spsi that enforces semigroup self consistency, guaranteeing that a single step inherits multi-step accuracy by ensuring
two consecutive small steps must agree with one large step.
This field absorbs the cumulative effect of marginal curvature, stabilizing the bootstrap signal during training and ensuring the d=1 limit matches the flow model vphi. -
Developed a dual mechanism where anchors suppress transport distance and spsi absorbs marginal curvature to address independent sources of single-step error. The prior
reduces mean transport distance by 36%,
while the displacement fieldabsorbs the cumulative effect of curvature in vphi via semigroup consistency.
Abstract
Generative motion prediction must satisfy three simultaneous requirements for real-world autonomy: high accuracy, diverse multimodal futures, and strictly bounded latency. Diffusion models meet the first two but violate the third, requiring tens to hundreds of denoising steps. We identify a conditioning strategy that resolves this tension: single-step integration is accurate when the underlying transport problem is local. A model that must both discover the correct behavioral mode and traverse a long displacement in one step accumulates large discretization errors; conditioning the base distribution to lie near plausible futures reduces the problem to short-range refinement, the regime where a single Euler step suffices. We instantiate this local transport conditioning in FlowS, a conditional flow matching framework with two mechanisms. First, an online, scene-conditioned learned prior emits K calibrated anchor trajectories per agent, each already near a plausible future, converting mode discovery into local correction. Second, a step-consistent displacement field enforces semigroup self-consistency, guaranteeing that a single step inherits multi-step accuracy. Crucially, anchoring this field at learned priors along straight-line paths yields a stable, low-variance training target, unlike prior self-consistency methods that suffer from high-variance bootstrap signals on curved diffusion paths. On the Waymo Open Motion Dataset, FlowS achieves state-of-the-art Soft mAP (0.4804) and mAP (0.4703) with ensemble at 75,FPS with single-step inference, demonstrating that local transport conditioning makes one-step generative motion prediction practical for safety-critical autonomy. Code and pretrained models will be released upon acceptance.
Sources
- Denoising Diffusion Bridge Models
- Flow Matching for Generative Modeling
- BridgeDrive: Diffusion Bridge Policy for Closed-Loop Trajectory Planning in Autonomous Driving
- Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow
- Improving and generalizing flow-based generative models with minibatch optimal transport
- TrajFlow: Multi-modal Motion Prediction via Flow Matching
- Wayformer: Motion Forecasting via Simple & Efficient Attention Networks
- FlowDrive: moderated flow matching with data balancing for trajectory planning
- One Step Diffusion via Shortcut Models
- FlowDrive: Energy Flow Field for End-to-End Autonomous Driving
- Consistency Flow Matching: Defining Straight Flows with Velocity Consistency
- IMPACT: Behavioral Intention-aware Multimodal Trajectory Prediction with Adaptive Context Trimming
Related papers
- Rethinking Actions: Continuity-Aware Priors for World-Model Planning
- MOSAIC: Modular Obstacle and Spatial-Aware Framework for Manipulation In Cluttered Scenes
- SAGAS: Semantic-Aware Graph-Assisted Stitching for Offline Temporal Logic Planning
- DexSeed: Expanding Human Hand-Object Interactions into Robot Demonstrations for Dexterous Policy Learning
- LaST0.5: A Generalist World Action Model with Interaction-Centric Latent Reasoning
- CIALLO: Learning Shared Intent Representations for Cross-Embodiment Manipulation