Mind the Noise: Sensitivity of Transformer-based Interaction-Aware Trajectory Prediction Models to Noisy Data

Researchers demonstrate that transformer-based trajectory prediction models used in autonomous vehicles experience severe accuracy degradation when exposed to noisy real-world sensor data, with prediction accuracy declining by up to 3.9x under realistic noise conditions. The findings highlight a critical gap between idealized training environments and actual deployment scenarios, signaling the need for robust noise mitigation strategies in autonomous vehicle systems.

The study addresses a fundamental vulnerability in modern autonomous vehicle perception systems: the disconnect between laboratory-validated models and real-world operational conditions. Transformer-based trajectory prediction models have achieved impressive accuracy rates on clean datasets, but this research exposes how quickly performance deteriorates when confronted with sensor noise and localization errors inherent in actual deployments. The degradation pattern—1.3x accuracy loss at low noise levels escalating to 3.9x at realistic thresholds—suggests non-linear failure modes that current training protocols fail to address.

This research emerges within a broader trend of AI safety validation. The autonomous vehicle industry has increasingly relied on high-quality, curated datasets for model development, creating a false confidence in real-world robustness. Vehicle-to-Everything (V2X) communications introduce additional uncertainty layers that laboratory conditions typically exclude. The gap between academic benchmarks and operational requirements represents a systemic challenge affecting multiple autonomous driving companies and perception algorithm developers.

For the autonomous vehicle industry, these findings carry substantial implications. Manufacturers and algorithm developers must fundamentally rethink validation methodologies, incorporating noise injection during training and developing explicit noise robustness evaluation metrics. The work suggests that current safety assurance frameworks may underestimate failure modes associated with sensor degradation. Companies relying on third-party perception stacks face particular risk if noise characteristics differ from training distributions.

Moving forward, the field should prioritize domain randomization techniques, uncertainty quantification in trajectory predictions, and development of noise-adaptive models. Regulatory bodies evaluating autonomous vehicle safety certification may need to mandate noise robustness testing before deployment approval.

• →Transformer trajectory prediction models experience 3.9x accuracy degradation under realistic sensor noise conditions
• →Current training datasets exclude detection and tracking noise that consistently appears in real-world autonomous vehicle deployments
• →Vehicle-to-Everything communication channels introduce perception uncertainties that standard evaluation protocols do not measure
• →Non-linear accuracy degradation patterns suggest current noise-handling approaches are fundamentally insufficient
• →Industry requires new validation frameworks incorporating realistic noise profiles before autonomous vehicle deployment