May the secret to safe autonomous driving not be in camera perfection, but instead in integrating them with other sensors? RoboSense founder and chief scientist Steven Qiu thinks so and he has the numbers to prove it. Speaking at the FutureChina Global Forum in Singapore, Qiu said bluntly, “By now, it is clear that everyone understands that a vision-only approach is not safe enough. There are a lot of corner cases that a vision-only system cannot account for.”
LiDAR Light Detection and Ranging operates on the principle of sending out laser pulses and calculating their time of return to create three-dimensional mappings of the environment. In current autonomous vehicles, LiDAR modules can provide more than 200,000 points of data per second, perform accurately under low-light scenarios, and sense objects at distances of more than 200 meters. These attributes make it extremely valuable for high-level automating, especially SAE Level 3 and Level 4 systems, where cars need to undertake sophisticated driving functions without constant human observation.
Tesla’s Full Self-Driving (FSD) system, which is currently rated at Level 2, uses just eight external cameras. This vision-only approach emulates human vision but has inherent shortcomings: depth perception is computed by the algorithm rather than measured directly, and performance suffers in bad lighting or weather. Qiu explained the issue using a concrete example: a vision-only system could find it difficult to tell apart a white parked car from an intense cloud reflection, or an approaching black car while driving into a tunnel.
Sensor fusion uniting data from various types of sensors is the technology of choice for addressing such problems. By using cameras, LiDAR, radar, and ultrasonic sensors together, autonomous systems can cross-check object identification, overcome the limitations of individual sensors, and provide continuous reliable perception across mixed environments. Multi-sensor fusion algorithms have been found to track objects accurately even in fog or heavy rain conditions when single-sensor solutions struggle.
Cost has long been Musk’s strongest argument against LiDAR. In 2019, he described it as “friggin’ stupid” for vehicles, citing the expense and redundancy once vision is optimized. As of that year, car-grade LiDAR could be as much as $12,000 per unit. But Qiu points out that prices have plummeted from about $70,000 per vehicle in early deployments to just a few hundred dollars today thanks to advances in manufacturing, economies of scale, and competition among suppliers. According to industry data, unit costs have dropped by roughly 90% since 2015, while performance has improved through higher resolution, faster refresh rates, and better environmental resilience.
Technical specifications of leading LiDAR systems underscore their growing appeal. Mechanical spinning LiDARs like the Velodyne VLS-128 provide a complete 360° horizontal field of view, vertical resolution as low as 0.11°, and detection ranges up to 245 meters. Solid-state LiDARs, while having a narrower field of view, provide durability and lower integration costs and thus become more suitable for mass-market vehicles. Such sensors emit at wavelengths such as 905 nm, which compromise between safety and atmospheric penetration, and are able to detect several returns per pulse to improve detection quality in heavy clutter.
Environmental conditions further underscore the gap between vision-only and multi-sensor solutions. Cameras are blinded by sun glare, hidden by darkness, or confused by reflective surfaces. LiDAR, independent of light, can identify objects accurately in these environments, though signal scatter might be encountered during heavy rain. Radar provides another layer of redundancy, penetrating fog and dust where cameras and LiDAR suffer loss of fidelity. Each of these modalities becomes part of a perception “stack” that is optimized for safety and reliability.
Traffic conditions also affect sensor selection. Li Xiang, the CEO of Li Auto, said that night driving in China typically involves coming across trucks parked on highways with no or defective taillights obstacles that camera-only systems would likely not detect at all. “I believe that if Musk were in China, and driving on various highways late in the night, he would choose to include LiDAR as well,” Li said.
The launch last week of Tesla’s camera-only robotaxi pilot in Austin, Texas, revealed the stakes of this fight. Footage at the launch depicted vehicles fighting to make simple maneuvers, incorrectly judging lanes, and suddenly braking problems that sensor fusion might solve. Regulators have noticed, with the National Highway Traffic Safety Administration asking Tesla to provide detailed incident data.
As the world moves towards increased autonomy, the engineering opinion is evolving. The “holy trinity” of cameras, radar, and LiDAR backed by sophisticated fusion algorithms is one way to deliver safer, more flexible autonomous systems. Qiu’s opinion is part of a wider acknowledgment that vision is important but not absolute, and the future of autonomous driving probably belongs to those who accept redundancy, accuracy, and robustness in sensor design.
