Abstract

The fundamental brittleness of trajectory prediction systems poses a critical challenge to the deployment of autonomous vehicles and robotic systems in dynamic real-world environments. While transformer-based models demonstrate exceptional performance on benchmark datasets, their static inference paradigm renders them vulnerable to distribution shifts—changes in environmental conditions, sensor characteristics, and motion patterns that inevitably arise during deployment. This limitation undermines reliability precisely when accurate predictions are most critical for safety-critical autonomous systems.

This thesis introduces RecTTA (Reconstruction-based Test-Time Adaptation), a groundbreaking framework that fundamentally transforms trajectory prediction from static inference to dynamic adaptation. Unlike existing domain-specific adaptation methods, RecTTA leverages input trajectory reconstruction as a universal self-supervised signal that naturally preserves the spatial-temporal dependencies essential for motion forecasting. Through joint training with an auxiliary reconstruction decoder, our approach enables transformer models to continuously refine their internal representations for each test sample without requiring ground truth supervision.

Our comprehensive evaluation on the JTA dataset reveals three paradigm-shifting discoveries. First, RecTTA achieves consistent and substantial performance improvements of 4.09\% ADE and 3.07\% FDE reduction across diverse scenarios, establishing reconstruction-based adaptation as a robust enhancement mechanism. Second, we make a counter-intuitive discovery that fundamentally challenges conventional adaptation wisdom: selective adaptation of only the final output layers dramatically outperforms full model adaptation (3.55\% vs 3.35\% ADE improvement) while requiring 64.0\% less computation time. This finding reveals that adaptation effectiveness concentrates in the prediction bottleneck rather than deep feature representations. Third, we demonstrate a democratizing effect where trajectory-only inputs achieve the largest relative improvements (7.50\% ADE, 9.57\% FDE), enabling resource-constrained systems to approach the performance of complex multi-modal configurations.

These contributions establish test-time adaptation as an essential paradigm for robust autonomous system deployment. By proving that strategic adaptation can simultaneously enhance performance and computational efficiency while democratizing access to advanced prediction capabilities, this work provides both theoretical insights and practical tools for the next generation of adaptive AI systems. The principles underlying RecTTA—reconstruction as universal adaptation signal, selective parameter optimization, and architecture-aware adaptation strategies—extend beyond trajectory prediction to establish foundational concepts for adaptive neural architectures in structured prediction tasks.