[Paper] Context-free Self-Conditioned GAN for Trajectory Forecasting

Source: arXiv - 2603.08658v1

Overview

A new paper introduces a self‑conditioned Generative Adversarial Network (GAN) that can forecast 2‑D trajectories without relying on any external context (e.g., map data, scene semantics). By letting the discriminator’s own feature space discover distinct motion “modes,” the model learns to predict multiple plausible future paths for humans and road agents in an entirely unsupervised way. The authors show that this context‑free approach can beat existing methods that need handcrafted labels or extra scene information.

Key Contributions

• Self‑conditioned GAN architecture that internally discovers motion modes in the discriminator’s latent space.
• Three training regimes (pure unsupervised, semi‑supervised self‑conditioning, and label‑aware self‑conditioning) that progressively improve forecasting quality.
• Mode‑aware loss that encourages the generator to produce diverse trajectories matching the discovered modes.
• Comprehensive evaluation on two benchmark datasets (human motion and road‑agent trajectories) demonstrating state‑of‑the‑art performance among context‑free methods.
• Evidence that unsupervised mode discovery can substitute for expensive manual labeling in trajectory prediction pipelines.

Methodology

1. Generator‑Discriminator Loop – As in a classic GAN, the generator proposes future trajectories given a past trajectory, while the discriminator tries to tell real from fake.
2. Self‑Conditioning Mechanism – After each discriminator forward pass, its intermediate feature vector is fed back as a conditioning signal for the next generator step. This creates a feedback loop where the discriminator’s own representation of “behavioral mode” guides the generator.
3. Mode Extraction – The discriminator’s feature space is clustered (e.g., via k‑means) on‑the‑fly, producing pseudo‑labels that represent distinct motion patterns (straight‑line, turning, stopping, etc.).
4. Training Settings
   • Pure Self‑Conditioned GAN (SC‑GAN) – No external labels; the model relies solely on the discovered modes.
   • Semi‑Supervised SC‑GAN – A small fraction of ground‑truth mode labels are injected to steer clustering.
   • Label‑Aware SC‑GAN – Full supervision is used only to evaluate the upper bound; the architecture remains unchanged.
5. Loss Functions – Standard adversarial loss plus a mode consistency loss that penalizes mismatches between the generator’s output and the discriminator’s current mode embedding.

The whole pipeline runs end‑to‑end on raw trajectory sequences, requiring only past positions as input.

Results & Findings

• The unsupervised SC‑GAN already outperforms the best published context‑free baselines on the least represented motion labels (e.g., rare turning maneuvers).
• Adding a tiny amount of labeled data (≤ 5 % of trajectories) yields further gains, closing the gap to fully supervised models.
• Qualitative visualizations show the model generating multiple plausible futures that respect the discovered modes, reducing mode collapse—a common GAN issue.

Practical Implications

• Reduced labeling cost – Developers can train robust trajectory forecasters without hand‑annotating behavior categories, which is especially valuable for new cities or domains where labeled data is scarce.
• Plug‑and‑play forecasting module – The self‑conditioned GAN can be dropped into autonomous‑driving stacks, crowd‑simulation engines, or AR/VR motion prediction pipelines, requiring only past positions as input.
• Diverse predictions out of the box – The mode‑aware design naturally yields a set of candidate trajectories, simplifying downstream decision‑making (e.g., risk assessment in autonomous vehicles).
• Scalable to edge devices – Because the model does not need heavy scene‑graph or map encoders, inference can be kept lightweight, making it suitable for on‑board deployment on robots or drones.
• Foundation for hybrid systems – The discovered modes can be combined with external context (road maps, semantic segmentation) to create a context‑augmented predictor that benefits from both unsupervised diversity and domain knowledge.

Limitations & Future Work

• Mode granularity depends on clustering – The number of modes is a hyper‑parameter; too few leads to under‑diversity, too many can fragment training. Adaptive clustering strategies are an open question.
• No explicit scene awareness – While the approach shines in context‑free settings, it may miss safety‑critical cues (e.g., traffic lights) that require environmental inputs.
• Evaluation limited to 2‑D trajectories – Extending the method to 3‑D motion (e.g., aerial drones) or to longer prediction horizons remains to be explored.
• Potential instability of GAN training – The authors note occasional mode collapse when the discriminator overfits; future work could integrate recent GAN stabilization tricks (e.g., spectral normalization, Wasserstein loss).

Overall, the paper demonstrates that self‑conditioning can turn a vanilla GAN into a powerful, label‑light trajectory forecaster, opening new avenues for developers who need flexible, data‑efficient prediction models.

Authors

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

Paper Information

• arXiv ID: 2603.08658v1
• Categories: cs.LG
• Published: March 9, 2026
• PDF: Download PDF