Context-free Self-Conditioned GAN for Trajectory Forecasting

Tiago Rodrigues de Almeida; Eduardo Gutierrez Maestro; Oscar Martinez Mozos

cs.LGMar 9, 2026

Context-free Self-Conditioned GAN for Trajectory Forecasting

Authors: Tiago Rodrigues de Almeida, Eduardo Gutierrez Maestro, Oscar Martinez Mozos

Organizations: AASS, Örebro University, Sweden

Abstract

In this paper, we present a context-free unsupervised approach based on a self-conditioned GAN to learn different modes from 2D trajectories. Our intuition is that each mode indicates a different behavioral moving pattern in the discriminator's feature space. We apply this approach to the problem of trajectory forecasting. We present three different training settings based on self-conditioned GAN, which produce better forecasters. We test our method in two data sets: human motion and road agents. Experimental results show that our approach outperforms previous context-free methods in the least representative supervised labels while performing well in the remaining labels. In addition, our approach outperforms globally in human motion, while performing well in road agents.

Figures & tables

Panel of Figure 1 — Fig. 1: Proposed framework’s overview. First, self-conditioned GAN learns the different modes of the input data. Then, this information is used as training settings (via soft-assumptions ) to improve the prediction in specific modes.

Panel of Figure 2 — Fig. 2: Self-conditioned GAN for identifying meaningful clusters, which represent different modes ( $\boldsymbol{m}$ ) of the input trajectory data. It comprises two branches: the discriminator ( $D$ ) and the generator ( $G$ ). During the former’s training (upper part), the inputs are fake or real samples, $\boldsymbol{{\hat{F}}}$ and $\boldsymbol{R}$ , respectively. Then, the encoder extracts features from those inputs. These features become semantically meaningful during training, so they are reclustered from time to time. Finally, a classifier produces the likelihood, $\boldsymbol{s}$ , of the respective input being a sample from the real distribution. The second branch (lower part) depicts the generator’s training. It takes as input the observed trajectories, $\boldsymbol{X}$ , concatenated with the respective modes, $\boldsymbol{m}$ . Before this, the clustering algorithm provides the respective modes based on the discriminator’s encoder features from the entire and real trajectory, $\boldsymbol{R}$ . Thereafter, the generator’s encoder extracts features from the input. Finally, these features plus the latent variable, $\boldsymbol{z}$ , are decoded, yielding the predicted trajectory, $\boldsymbol{{\hat{Y}}}$ .

		Baselines		Ours
Data set	Labels (# samples)	LSTM [ 22 ]	CF VAN GAN [ 14 ]	CF VAN GAN + wL2	CF VAN GAN + wB	CF VAN GAN +wL2 + wB
THÖR	workers (413)	0.695 1.064	$0.642\pm 0.006$ $1.033\pm 0.005$	$0.629\pm 0.005$ $1.009\pm 0.014$	$0.644\pm 0.012$ $1.044\pm 0.028$	$\mathbf{0.625\pm 0.009}$ $\mathbf{1.006\pm 0.019}$
	visitors (1379)	0.664 1.139	$0.660\pm 0.001$ $\mathbf{1.105\pm 0.090}$	$\mathbf{0.657\pm 0.003}$ $1.107\pm 0.007$	$0.668\pm 0.005$ $1.124\pm 0.018$	$0.657\pm 0.003$ $1.113\pm 0.013$
	inspector (260)	0.796 1.582	$0.735\pm 0.007$ $1.474\pm 0.019$	$0.736\pm 0.008$ $\mathbf{1.473\pm 0.013}$	$\mathbf{0.729\pm 0.013}$ $1.479\pm 0.049$	$0.734\pm 0.003$ $1.476\pm 0.015$
Argoverse	others (526)	1.864 3.029	$1.815\pm 0.031$ $2.969\pm 0.034$	$1.799\pm 0.007$ $2.944\pm 0.022$	$\mathbf{1.789\pm 0.012}$ $2.927\pm 0.020$	$1.801\pm 0.027$ $\mathbf{2.919\pm 0.032}$
	av (2600)	1.512 2.278	$\mathbf{1.467\pm 0.007}$ $\mathbf{2.269\pm 0.023}$	$1.482\pm 0.009$ $2.292\pm 0.010$	$1.480\pm 0.003$ $2.282\pm 0.006$	$1.493\pm 0.010$ $2.298\pm 0.028$

TABLE I: Intra-classes ADE/FDE metrics (in meters) in the test sets.

		Baselines		Ours
Data set	Cluster ID (# samples)	LSTM [ 22 ]	CF VAN GAN [ 14 ]	CF VAN GAN + wL2	CF VAN GAN + wB	CF VAN GAN +wL2 + wB
THÖR	9 (23)	1.203 2.456	$1.120\pm 0.025$ $2.758\pm 0.082$	$1.054\pm 0.048$ $2.505\pm 0.134$	$1.124\pm 0.038$ $2.811\pm 0.136$	$\mathbf{1.039\pm 0.055}$ $\mathbf{2.505\pm 0.105}$
THÖR	0 (1003)	0.325 0.447	$\mathbf{0.311\pm 0.003}$ $\mathbf{0.403\pm 0.010}$	$0.321\pm 0.004$ $0.419\pm 0.015$	$0.317\pm 0.007$ $0.416\pm 0.021$	$0.315\pm 0.002$ $0.424\pm 0.017$
Argoverse	10 (16)	7.394 19.075	$7.184\pm 0.178$ $18.402\pm 0.415$	$7.105\pm 0.123$ $18.233\pm 0.297$	$7.122\pm 0.055$ $18.276\pm 0.113$	$\mathbf{7.047\pm 0.088}$ $\mathbf{18.128\pm 0.194}$
	18 (1542)	0.912 1.148	$0.809\pm 0.016$ $1.100\pm 0.017$	$0.807\pm 0.010$ $1.088\pm 0.027$	$0.805\pm 0.007$ $1.079\pm 0.012$	$\mathbf{0.795\pm 0.008}$ $\mathbf{1.055\pm 0.030}$

TABLE II: ADE/FDE metrics (in meters) for 2 clusters of the test set.

	Baselines		Ours
Data set	LSTM [ 22 ]	CF VAN GAN [ 14 ]	CF VAN GAN + wL2	CF VAN GAN + wB	CF VAN GAN +wL2 + wB
THÖR	$0.685$ $1.163$	$0.663\pm 0.002$ $1.123\pm 0.019$	$\mathbf{0.658\pm 0.003}$ $\mathbf{1.119\pm 0.009}$	$0.668\pm 0.003$ $\mathbf{1.119\pm 0.009}$	$\mathbf{0.658\pm 0.002}$ $1.122\pm 0.013$
Argoverse	$1.948$ $3.330$	$\mathbf{1.912\pm 0.012}$ $\mathbf{3.298\pm 0.014}$	$1.915\pm 0.012$ $3.307\pm 0.017$	$1.915\pm 0.009$ $3.310\pm 0.018$	$1.923\pm 0.010$ $3.307\pm 0.014$

TABLE III: ADE/FDE metrics (in meters) in the test sets.

Data set	cGAN	Ours
THÖR	$0.657\pm 0.003$ $1.114\pm 0.011$	$\mathbf{0.591\pm 0.014}$ $\mathbf{0.937\pm 0.022}$
Argoverse	$1.927\pm 0.016$ $3.323\pm 0.030$	$\mathbf{1.785\pm 0.014}$ $\mathbf{2.887\pm 0.038}$

TABLE IV: ADE/FDE metrics (in meters) in the test sets for ideal models. These can be seen as lower bounds.

Fig. 3: Example of trajectory forecasting in the THÖR (left) and the Argoverse (right) data sets for complex trajectories, being

X

the observed track and

Y

and

\hat{Y}

the ground truth and the predictions, respectively. We can see how our approach gets closer to the ground truth in these cases.

Fig. 4: Entire trajectories randomly sampled from THÖR (left) and Argoverse (right) data sets, where crosses mark the starting point of each track. In THÖR, trajectories from cluster 9 go from right to left, while tracks from cluster 10 go from left to right. In Argoverse, trajectories from cluster 0 are much longer than the ones from cluster 18.

Fig. 5: Test set examples of trajectories generated from the self-conditioned GAN with four different labels in both data sets (THÖR on the left and Argoverse on the right). Both cases show that the right condition (represented by

\hat{Y}_{Ours}

) yields trajectories closer to the ground truth when compared to other three different modes (

m_{1}

m_{2}

, and

m_{3}

Trajectory Prediction

Trajectory prediction forecasts how people, vehicles, robots, and other dynamic systems will move from observed histories and context.

See all 187 papers

Oct 5, 2026

MoCAR: Motion-code Coordinate-aware AutoRegression for Continuous Trajectory Forecasting

Autoregressive generation is natural for language, where predicted tokens can be directly reused as the next prediction state, but trajectory forecasting lacks such a clean token: motion is continuous, multimodal, and expressed in local coordinate frames that evolve with the predicted trajectory. We present MoCAR (Motion-code Coordinate-aware AutoRegression), a decoder-only framework that casts trajectory forecasting as next-code prediction in a coordinate-aware continuous latent space. MoCAR learns a continuous motion-code space from endpoint-normalized trajectory segments, where each code jointly captures local trajectory geometry and the reference-frame transition induced by that segment. Historical motion codes are used as a teacher-forced prefix, future codes are generated autoregressively under temporal, map, agent, and mode interactions, and predicted codes persist in latent memory while decoded endpoints update the local scene context. This enables rollout without trajectory-space re-tokenization, trajectory queries, goal candidates, or proposal-and-refinement pipelines. On Argoverse (AV) benchmarks, MoCAR achieves top-tier performance with a simple single-stage architecture, transfers strongly from AV2 to AV1 in zero-shot evaluation, and improves on turn-heavy scenarios. Ablations confirm that the learned continuous motion-code space, latent alignment, weak KL regularization, and joint tokenizer-predictor optimization are essential for stable latent autoregression.

Yiming Xu, Hao Cheng, Monika Sester

Institute of Cartography and Geoinformatics Leibniz University Hannover, 30167 Hannover, Germany · Faculty of Geo-Information Science and Earth Observation University of Twente, 7522 NH Enschede, The Netherlands

Trajectory Generation Trajectory Representation Learning Multimodal Trajectory Prediction Trajectory Prediction

Ask PDF

Jun 24, 2026

Rethinking Training & Inference for Forecasting: Linking Winner-Take-All back to GMMs

Trajectory forecasting for autonomous driving has advanced rapidly, yet representative models often produce uninformative posteriors over forecast modes, causing problems for mode pruning. We trace this to a modeling-training mismatch: forecasters are typically modeled as conditional Gaussian mixture models (GMMs) but trained with a winner-take-all (WTA) loss that assigns each sample to its nearest mode. We argue that this K-means-like hard assignment (one-hot), while preventing mode collapse, is the source of uninformative mode probabilities: it over-segments the trajectory space, ignores relatedness among nearby modes, and yields assignment instability under small perturbations. Guided by this lens, we introduce two post-hoc treatments: (1) test-time posterior-weighted merging that aggregates nearby candidate trajectories; and (2) a one-step expectation-maximization (EM) update that replaces hard labels with soft responsibilities, sharing probability mass across neighboring modes. Across several WTA-trained architectures, these lightweight steps produce more informative, faithfully ranked mode posteriors and strengthen final forecasts on popular displacement metrics -- without retraining. Our analysis unifies recent design choices through a GMM-vs-K-means perspective and offers principled, practical corrections that better align training objectives with inference.

Qiyuan Wu, Katie Z Luo, Bharath Hariharan +2

Cornell University · Stanford University · Boston University

Gaussian Mixture Models Probabilistic Forecasting Multimodal Trajectory Prediction Vehicle Trajectory Prediction

Ask PDF

Apr 28, 2026

FlowS: One-Step Motion Prediction via Local Transport Conditioning

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.

Leandro Di Bella, Adrian Munteanu, Bruno Cornelis

Department of Electronics and Informatics, Vrije Universiteit Brussel, Pleinlaan 2, B-1050 Brussels, Belgium · IMEC, Kapeldreef 75, B-3001 Leuven, Belgium

Flow Matching Trajectory Generation Multi-Agent Trajectory Prediction Autonomous Driving Perception One-Step Generative Modeling Flow-Based Generative Modeling+2

Ask PDF

Abstract

Trajectory Prediction

Explore similar work

MoCAR: Motion-code Coordinate-aware AutoRegression for Continuous Trajectory Forecasting

Rethinking Training & Inference for Forecasting: Linking Winner-Take-All back to GMMs

FlowS: One-Step Motion Prediction via Local Transport Conditioning