LiDAR vs. Camera-Only: The Great Self-Driving Sensor Debate
Few debates in the mobility industry generate more heat — and less consensus — than the sensor architecture for self-driving vehicles. On one side, Tesla and a growing cohort of Chinese startups insist that cameras alone, powered by sufficiently capable neural networks, can achieve full autonomy. On the other, Waymo, Cruise, and most European OEMs maintain that LiDAR is indispensable for safety-critical perception. At Güil Mobility Ventures, we have spent considerable time evaluating both approaches, and the answer is more nuanced than either camp admits.
The camera-only thesis
Tesla’s position, championed by Elon Musk with characteristic conviction, rests on a compelling logic: humans drive using only vision, so a sufficiently advanced vision system should be able to do the same. Tesla’s FSD (Full Self-Driving) stack processes feeds from eight cameras using a custom neural network running on the company’s HW4 inference chip. The approach leverages Tesla’s unmatched fleet data advantage — billions of miles of real-world driving footage used to train increasingly capable models.
The economic argument is equally powerful. A camera module costs roughly $10–$50 per unit. A full eight-camera suite adds perhaps $200–$400 to vehicle cost. This makes vision-based autonomy scalable across every price segment, from the Model 3 to a hypothetical $25,000 compact vehicle.
The limitations are real, however. Cameras struggle in adverse lighting conditions — direct sun glare, heavy rain, and snow significantly degrade image quality. Depth estimation from monocular or stereo cameras, while improving rapidly with transformer-based architectures, introduces measurement uncertainty that LiDAR simply does not have.
The LiDAR thesis
Waymo’s sensor suite illustrates the opposite philosophy. Each Waymo Jaguar I-PACE carries five LiDAR units, six radar units, and 29 cameras, creating a dense 360-degree perception bubble with centimeter-level depth accuracy. LiDAR’s time-of-flight measurement provides ground-truth range data regardless of lighting conditions — it works identically at noon and midnight, in sunshine and fog.
The cost objection that long defined the anti-LiDAR argument is rapidly fading. Luminar’s Iris LiDAR unit is priced under $1,000 at volume, and Chinese manufacturers like Hesai and RoboSense have pushed automotive-grade units below $500. At these price points, LiDAR adds modest cost to a premium vehicle while providing a fundamentally different — and arguably more reliable — perception modality.
The redundancy argument is perhaps the strongest case for LiDAR inclusion. In safety-critical systems, independent sensor modalities that fail in uncorrelated ways provide a stronger safety envelope than any single modality, no matter how capable. This is a foundational principle in aerospace and industrial safety engineering.
The convergence we see emerging
Our view at Güil is that the industry is converging toward a tiered approach. Robotaxi fleets and premium autonomous vehicles will use full sensor suites including LiDAR, radar, and cameras — the cost is justified by the commercial model and the safety requirements. Consumer ADAS and supervised autonomy features will increasingly rely on camera-centric architectures with radar as a secondary modality, because the cost-per-vehicle economics demand it.
The most interesting investment opportunities, from our perspective, sit at the intersection: companies building perception software that can fuse data from any combination of sensors, enabling OEMs to scale their autonomy stack across vehicle segments without rewriting the perception pipeline for each sensor configuration.
The investor takeaway
Sensor hardware is commoditizing. The enduring value lies in the software that interprets sensor data, the simulation platforms that validate perception performance, and the data infrastructure that enables continuous learning. We advise our portfolio companies to build sensor-agnostic architectures and let the hardware debate resolve itself through market forces.