Title: DeepIPCv2: Leveraging LiDAR for Resilient Perception and Navigation in Autonomous Driving

Abstract:

This paper introduces DeepIPCv2, a novel end-to-end autonomous driving system that unifies LiDAR-driven environmental understanding with control learning tailored to specific commands. Departing from previous models that rely heavily on cameras, DeepIPCv2 utilizes point cloud segmentation and multi-view projection techniques to build resilient scene representations. These representations are processed and decoded using a hybrid architecture comprising gated recurrent units, command-specific multi-layer perceptrons, and PID controllers, enabling the simultaneous estimation of waypoints and navigational commands. This architectural choice improves vehicle maneuverability and mitigates action imbalance commonly found in driving datasets.

To evaluate the framework, we developed a dataset encompassing various lighting conditions and performed both ablation studies and comparative analyses against contemporary methods, such as TransFuser. The results indicate that DeepIPCv2 yields the lowest total metric error and requires the fewest driving interventions, demonstrating superior robustness to lighting variations and enhanced control precision. We plan to open-source the code at https://github.com/oskarnatan/DeepIPCv2 to foster reproducibility and advance research in end-to-end autonomous driving.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC