How do autonomous driving systems work in modern electric vehicles? What types of sensors (LiDAR, radar, cameras) are used to build a 3D map of the surroundings, and how do machine learning algorithms process this data to make decisions? What are the main technical and regulatory obstacles currently standing in the way of full autonomy? I’d love to hear opinions on which approaches to integrating these technologies seem most promising.
Self-driving electric cars: what technologies make them possible?
👁️ 227 views💬 1 replies❤️ 0 likes
1 Replies
Currently, the autopilot in electric vehicles consists of multiple levels of sensors: LiDAR—the most precise depth source but expensive and weather-sensitive; radar—robust in harsh conditions but with limited angular resolution; and cameras—the cheap "eyes-on" module providing color and context but requiring complex algorithms for object recognition. In practice, most manufacturers combine all three types into a "sensor suite" to compensate for each individual sensor's weaknesses.
The fused data is fed into neural networks, where initial classification (pedestrian, vehicle, road sign) and semantic 3D mapping occur. Next in the pipeline are "trajectory planners," which use reinforcement learning or MPC approaches to simulate multiple hypothetical scenarios and select the optimal one considering safety and comfort. The issue is that models are trained on limited real-world road data, and the real world always has a "corner case" where the system may fail.
Current technical obstacles primarily include scaling high-performance chips to process terabytes of data in real time and sensor reliability in extreme conditions (rain, snow, fog). Regulatory barriers include the lack of a unified international "Level 5" standard, as well as strict verification and cybersecurity requirements that significantly delay the release of fully autonomous systems.
From my perspective, a more realistic path is to focus on "data-centric" processing, where LiDAR is used only in critical scenarios (e.g., merging onto a highway), and the rest of perception relies on cameras and radars backed by edge-ML accelerators. This hybrid approach reduces costs while maintaining sufficient reliability. What are your thoughts on reducing LiDAR dependency in near-term models?