Answers are grounded in this paper and cite their sources.
Evaluates a new hybrid perception-control approach in autonomous driving, suggesting improved safety and reliability.
End-to-end autonomous driving maps perception inputs directly into control commands. Although this paradigm is efficient, it remains challenging under complex interactions and long-horizon driving scenarios. A single perception encoder may fail to capture both local details and global context, which weakens robustness in recognizing small objects and modeling interaction semantics. In addition, control generation may not fully exploit planning cues, leading to delayed or inconsistent responses in sudden conflict scenarios. This paper proposes HPTC, a hybrid perception–trajectory–control framework for end-to-end autonomous driving. HPTC employs a Hybrid Perception Encoder that combines CNN-based local feature extraction with Transformer-based global context modeling. It also introduces a Trajectory–Control Attention module that injects multi-step trajectory information into control generation and improves decision stability. To handle long-tail, zerotolerance rules such as stop signs, HPTC further incorporates an inference-time rule intervention mechanism without changing the end-to-end training pipeline. Experiments on CARLA show that HPTC achieves route completion rates of 96.80% on Town05 Long and 84.50% on Longest6, with driving scores of 63.80 and 60.30, respectively, using only RGB input. Fine-grained analysis shows that the rule intervention mechanism reduces stop-sign violations from 0.47 to 0.01 while keeping other infraction categories largely stable.These results support the development of safer and more reliable automated-driving components for intelligent transportation systems.
No takes yet. Share an insight, caveat, or question.
Xiong et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: