DeepIPCv2: LiDAR-powered Robust Environmental Perception and Navigational Control for Autonomous Vehicle

DeepIPCv2 is an end-to-end autonomous driving framework that uses LiDAR point cloud data instead of cameras to perceive environments and control vehicle navigation. The system demonstrates superior robustness to lighting variations and reduced driving interventions compared to existing methods like TransFuser, advancing the practical deployment of autonomous vehicles.

DeepIPCv2 represents a significant methodological shift in autonomous driving research by prioritizing LiDAR-based perception over camera-centric approaches. This architectural choice addresses a fundamental limitation in existing systems: camera performance degrades dramatically under poor lighting conditions, fog, and glare. By leveraging point cloud segmentation and multi-view projection, the framework builds environmental representations independent of lighting conditions—a critical requirement for real-world autonomous vehicle deployment across diverse geographies and times of day.

The research landscape for autonomous driving has historically oscillated between perception modalities, with recent trends favoring camera-based vision systems due to computational efficiency and cost. However, this work demonstrates that sensor fusion strategies combining LiDAR with machine learning can achieve superior performance metrics. The integration of gated recurrent units for temporal reasoning, command-specific MLPs for specialized control outputs, and PID controllers for steering precision creates a hybrid classical-learning architecture that balances interpretability with performance.

For the autonomous vehicle industry, this advancement has immediate implications for safety-critical applications. Reduced driving interventions indicate fewer safety-critical situations requiring human takeover, directly improving the commercial viability of autonomous fleets. The comprehensive testing across illumination conditions validates real-world applicability beyond controlled environments. The authors' commitment to open-sourcing the code accelerates industry adoption and enables rapid validation by competitors and regulatory bodies.

• →LiDAR-based perception outperforms camera systems in varying illumination conditions, addressing a critical safety gap
• →DeepIPCv2 achieved lowest total metric error and fewest driving interventions compared to recent autonomous driving methods
• →Hybrid approach combining deep learning with classical control (PID) improves both robustness and interpretability
• →Open-source release supports reproducibility and industry-wide advancement in end-to-end autonomous driving
• →Multi-modal sensor fusion strategy represents shifting paradigm away from camera-only autonomous vehicle architectures