acceptodds
Under review as a conference paper at ICLR 2027

TrajCFM: Observation-Conditioned Flow Matching for Unified Trajectory Modeling

Abstract

Real-world trajectories combine irregular sampling, partial observability, and heterogeneous motion dynamics. Modeling them across tasks requires capturing local motion structure under different observation conditions. We present TrajCFM, an observation-conditioned flow matching framework for unified trajectory modeling. A DiT-style Transformer, conditioned on flow time and observed trajectory context, learns a velocity field that transports noise at unobserved positions toward target coordinates while preserving observed locations. This formulation brings completion and prediction under a shared objective and backbone. To capture motion at varying temporal scales, observation-aware adaptive patching builds variable-length segment representations from local motion cues and observation availability. Sparse mixture-of-experts routing selectively activates experts to model heterogeneous trajectory patterns. Pretraining with diverse observation masks learns representations that can also be adapted to trajectory classification and road-conditioned generation. Experiments on five real-world datasets demonstrate that TrajCFM achieves state-of-the-art performance in both trajectory completion and prediction using only three flow integration steps. Downstream experiments on transport-mode classification across GeoLife and Grab-Posisi, together with road-conditioned generation, demonstrate the transferability of the learned representations.

Primary area:
learning on time series and dynamical systems
Keywords:
Trajectory modeling, Conditional flow matching, Adaptive patching, Mixture of experts, Trajectory completion, Trajectory prediction, Representation learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_eGvpOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes TrajCFM, an observation-conditioned flow-matching framework that unifies trajectory completion and prediction under different observation masks. It combines a DiT-style Transformer with observation-aware adaptive patching and mixture-of-experts routing to model irregular and heterogeneous trajectory dynamics. Experiments on five real-world datasets report good performance on completion and prediction, with additional transfer results on trajectory classification and road-conditioned generation.

Critical strengths

S1. The unified treatment of completion and prediction is practically useful and the framework supports additional downstream transfer tasks.

S2. The empirical results are good across multiple real-world trajectory datasets and tasks.

Critical weaknesses

W1. Novelty is limited. The framework largely combines established ingredients: conditional flow matching, a DiT-style Transformer, adaptive patching, mixture-of-experts routing, and masked/observation-conditioned training, rather than introducing a clearly new modeling principle.

W2. The experimental comparison is incomplete for closely related unified trajectory models. GenMove and UniMove are discussed but not empirically compared, and MoFlow (CVPR 2025) is also highly relevant as a recent flow-matching trajectory model.

Questions for authors

Q1. Please clarify the core technical novelty of TrajCFM relative to prior unified trajectory modeling and conditional flow-matching approaches. Which component or formulation is fundamentally new rather than a combination of existing techniques?

Q2. Can the authors add empirical comparisons with the closely related methods already discussed in the paper, particularly GenMove, UniTraj, UniMove, as well as MoFlow? If a direct comparison is not feasible, please explain the specific incompatibility or evaluation mismatch.

Reviewer_Zq4QOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes TrajCFM, an observation-conditioned flow matching framework for unified trajectory modeling. It formulates trajectory completion and prediction as a shared conditional reconstruction task by employing different observation masks. Experiments on five real-world datasets demonstrate competitive performance in trajectory completion and prediction, with additional evaluations on downstream tasks.

Critical strengths

(1)The paper provides a unified formulation for trajectory completion and prediction through observation-conditioned flow matching, allowing both tasks to share the same network architecture and training objective.

(2)The proposed observation-aware adaptive patching combines point-level and segment-level representations, allowing the model to incorporate local contextual information while preserving coordinate-level resolution.

(3)The method is evaluated on five real-world datasets and achieves strong performance in both trajectory completion and prediction.

Critical weaknesses

(1) The three main components of the proposed framework, conditional Flow Matching, adaptive patching, and Sparse MoE. are all based on relatively well-established techniques. Consequently, the core contribution appears to be primarily an architectural combination and engineering-oriented optimization tailored to trajectory data, rather than a fundamentally novel methodological advancement.

(2) The paper claims that Sparse MoE can select appropriate computational experts for different motion patterns. However, Figure 2 primarily illustrates the distribution of expert routing weights across different motion groups. Differences in expert utilization do not necessarily indicate that the experts have learned distinct motion dynamics.

(3) Although the paper discusses generative trajectory and time-series models such as MoFlow, GenMove, TrajFlow, and CSDI in the related work, the main experiments primarily compare against Linear, DHTR, Transformer, DeepMove, TrajBERT, TrajFM, and UniTraj. For a paper centered on Flow Matching, the absence of direct comparisons with conditional diffusion models, existing trajectory Flow Matching methods, and other strong generative baselines represents a significant limitation.

(4) The efficiency evaluation lacks comprehensive comparisons with UniTraj, TrajFM, and other strong baselines in terms of inference latency, FLOPs, GPU memory consumption, and parameter count under identical hardware and evaluation settings.

Questions for authors

See weaknesses

Reviewer_F4d3OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes TrajCFM, an observation-conditioned flow matching framework for unified trajectory modeling. It combines a DiT-based backbone with adaptive patching and sparse mixture-of-experts to support trajectory completion and prediction under different observation masks. The learned representations are further adapted to trajectory classification and generation. Experiments are conducted on several real-world trajectory datasets.

Critical strengths
  1. Unified framework for trajectory reconstruction. The paper provides a unified formulation for trajectory completion and prediction through different observation masks, allowing both tasks to share the same backbone and inference procedure.

  2. Reasonable architectural evaluation. The paper includes ablation studies and additional controlled experiments to investigate the contributions of adaptive patching and MoE, providing some evidence of their effectiveness within the proposed architecture.

Critical weaknesses
  1. Unfair experimental comparisons and inconsistent evaluation protocols. The main experimental comparisons in Tables 1 and 2 appear to directly reuse baseline results from the UniTraj paper [1], despite differences in the following experimental protocols: i) The reported UniTraj results correspond to its zero-shot setting, while TrajCFM is fine-tuned on target datasets. The substantially stronger fine-tuned UniTraj results are not included. ii) The evaluation protocols are inconsistent. UniTraj evaluates on 100,000 WorldTrace test trajectories, whereas this paper reports 245,129 test trajectories. These issues fundamentally undermine the fairness of the comparisons and the claimed state-of-the-art performance. Without evaluations under consistent protocols, the main experimental conclusions are not convincing.

  2. Significantly limited methodological novelty. The proposed method largely combines existing techniques, and its core components have been extensively explored in prior top-tier conference publications: UniTraj [1] explores unified trajectory representation learning through masked pretraining. Diffusion-TS [2] supports time-series imputation and forecasting within a unified generative framework. MoFlow [3] applies conditional flow matching to trajectory forecasting. Pathformer [4] explores adaptive multi-scale patch-based modeling. Time-MoE [5] employs sparse expert routing for time-series modeling. TrajCFM essentially integrates these established ideas into a DiT-based architecture. Although the specific implementation differs, the paper does not introduce a sufficiently novel modeling principle or demonstrate insights beyond combining existing techniques.

[1] UniTraj: Learning a Universal Trajectory Foundation Model from Billion-Scale Worldwide Traces NeurIPS 2025. [2] Diffusion-TS: Interpretable Diffusion for General Time Series Generation ICLR 2024. [3] MoFlow: One-Step Flow Matching for Human Trajectory Forecasting via Implicit Maximum Likelihood Estimation based Distillation CVPR 2025. [4] Pathformer: Multi-scale Transformers with Adaptive Pathways for Time Series Forecasting ICLR 2024. [5] Time-MoE: Billion-Scale Time Series Foundation Models with Mixture of Experts ICLR 2025.

Questions for authors
  1. Why are the published zero-shot UniTraj results used for comparison when TrajCFM is fine-tuned on the target datasets? How can the comparisons be considered fair given the differences in evaluation protocols?
  2. What is the fundamental methodological novelty of TrajCFM beyond combining existing mask-conditioned modeling, flow matching, adaptive patching, and MoE techniques?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.