UniAd
2 cited papers · May 27, 2026 · Powered by Researchly AI
The query "UniAD" most likely refers to a unified autonomous driving framework. However, based on the retrieved evidence, the closest relevant work addresses ad…
- End-to-End Autonomous Driving (AD) Models — Systems that integrate perception, prediction, and planning into a unified pipeline, achieving state-of-the-art performance but remaining vulnerable to adversarial perturbations. Zhang et al. (2024)
- Adversarial Perturbations — Human-imperceptible noise-like patterns that can subtly alter input data and disrupt decision-making in otherwise accurate deep learning models.
- Module-wise Noise Injection — A technique that injects noise before the input of different modules, targeting training with the guidance of overall objectives rather than each independent module loss.
- Dynamic Weight Accumulation Adaptation — A component of MA2T designed to address the strongly interconnected and distinct objectives of different stages within end-to-end AD models.
End-to-End Autonomous Driving Pipeline (MA2T Framework) ┌─────────────────────────────────────────────────────┐ │ Input (Camera/Sensor) │ └───────────────────────┬─────────────────────────────┘ │ ┌─────────────▼──────────────┐ │ Module-wise Noise │ │ Injection Layer │ └─────────────┬──────────────┘ │ ┌───────────────┼───────────────┐ ▼ ▼ ▼ [Perception] [Prediction] [Planning] Module Module Module │ │ │ └───────────────┼───────────────┘ │ ┌─────────────▼──────────────┐ │ Dynamic Weight │ │ Accumulation Adaptation │ └─────────────┬──────────────┘ │ ┌─────────────▼──────────────┐ │ Overall Objective Loss │ │ (Adversarial Training) │ └────────────────────────────┘
MA2T addresses a critical gap: no prior studies had focused on adversarial training for end-to-end AD models before this work. The challenge is non-trivial because different stages within the model have distinct objectives and are strongly interconnected, making conventional adversarial training extensions difficult.
| Aspect | Conventional Adversarial Training | MA2T (End-to-End AD) |
|---|---|---|
| Target Domain | Image/text classifiers | Autonomous driving pipeline |
| Noise Injection | Input-level | Module-wise |
| Loss Guidance | Per-module loss | Overall objective |
| Interconnection Handling | Not addressed | Dynamic Weight Adaptation |
The retrieved evidence does not provide benchmark numbers (e.g., nuScenes scores) for MA2T or UniAD specifically, limiting quantitative comparison. Furthermore, extending adversarial training to multi-stage pipelines is described as "highly non-trivial," suggesting that the proposed solutions may not fully resolve all inter-module dependency challenges.
- End-to-end AD models integrating perception, prediction, and planning remain vulnerable to human-imperceptible adversarial perturbations.
- MA2T is presented as the first adversarial training framework specifically designed for end-to-end autonomous driving models.
- Adversarial perturbations are a cross-domain threat, affecting vision, text, graph, and speech models alike.
- The independence of defense mechanisms from target model structure is a desirable property, as demonstrated in image-domain defenses using conditional GANs.
- "UniAD unified autonomous driving perception prediction planning nuScenes benchmark"
- "Adversarial robustness evaluation end-to-end autonomous driving models empirical results"
- "Module-wise adversarial training multi-task deep learning robustness"
Research smarter with AI-powered citations
Researchly finds and cites academic papers for any research topic in seconds. Used by students across India.