Over the past few years, the most debated topic in the autonomous driving industry has been whether LiDAR is necessary. However, by 2026, the debate should shift to: what more can LiDAR do?Recently, the LiDAR industry has released a succession of reports: chips replacing optical stacking, point clouds acquiring color, and the number of lines in a single LiDAR unit jumping from over a hundred to nearly a thousand. Standing in 2026 and looking back at the LiDAR industry, what technological changes have occurred in this field?
What Changes Are Occurring in Perception Dimensions?
The primary task of LiDAR is to measure distance. By emitting laser pulses and calculating the reflection time, it generates a 3D point cloud. Each point in this coordinate system carries X, Y, and Z coordinates but lacks color information. Therefore, for an autonomous driving system to understand what objects are ahead, it requires algorithm-level fusion of the LiDAR point cloud with camera images. This process necessitates complex calibration and faces challenges in time synchronization.
Hesai Technology's Picasso SPAD-SoC chip, released on April 17, 2026, offers a different solution. It natively fuses RGB color sensing and ToF ranging at the pixel level on the chip hardware, directly outputting colored point clouds. The photon detection efficiency of the Picasso chip has surpassed 40%, reaching an internationally leading level (Related reading: Can LiDAR also perform color recognition?).
Image source: Hesai Technology
Photon detection efficiency is the core metric for evaluating the sensing capability of SPAD chips, directly determining how far and how clearly the LiDAR can see. Under the same transmission power, chips with higher photon detection efficiency can detect targets at longer distances and smaller sizes, and see more clearly in darker environments. The ETX series LiDAR equipped with this chip supports 1080-line, 2160-line, and 4320-line full-color 4K ultra-high-definition perception, with a maximum detection range of 600 meters, and 400 meters at 10% reflectivity. It can clearly identify 120x60 cm water-filled barriers within 300 meters, small animals within 280 meters, and 15x25 cm small wooden blocks within 150 meters.
This system no longer requires backend color fitting; the perception model can directly identify semantic information such as traffic lights and lane lines from the point cloud. In addition, the Picasso chip fully supports Hesai's self-developed waveform decoding engine and coded anti-interference technology, reducing the probability of false alarms and missed detections.
Besides enabling LiDAR to see colors, the number of lines in LiDAR is becoming increasingly competitive in 2026. The so-called number of lines refers to the number of vertical scanning channels of the LiDAR. The more channels there are, the denser the point cloud, and theoretically, the higher the perception fineness. In the past few years, the industry generally believed that 128 lines and 192 lines were sufficient, but in 2026, a higher number of lines is becoming a new technological threshold.
Huawei's 896-line dual-optical-path image-grade LiDAR, released on March 4, 2026, is currently the highest-specification mass-produced automotive LiDAR product globally. Unlike traditional LiDAR that uses a single set of receiving units, Huawei integrates two sets of receiving units inside the radar. The wide-angle unit covers a 120° field of view (FOV), responsible for multi-lane urban areas and complex blind spots; the telephoto unit focuses on a 30° FOV, responsible for fine long-distance recognition at high speeds. The two units can be controlled independently, and the algorithm will automatically adapt to different scenarios, such as turning on only the wide-angle or both wide-angle and telephoto simultaneously.
Compared with the previously launched 192-line LiDAR by HIMA (Harmony Intelligent Mobility Alliance), the new generation 896-line LiDAR has improved its imaging resolution by four times. Real-world test data shows that the minimum identifiable target height has dropped from 30 cm to 14 cm, close to the ground clearance of a typical sedan. The maximum recognition distance for 30 cm high obstacles has increased from 100 meters to 162 meters; in pitch-dark night environments, the maximum recognition distance for low-reflectivity targets such as tires has increased from 42 meters to 122 meters.
The single-frame point cloud volume of the 896-line dual-optical-path architecture is equivalent to 7 times that of 128 lines. The perception and recognition distance for low-reflectivity targets such as fallen tires has increased by 190%, and for irregular obstacles such as horizontally fallen traffic cones, it has increased by 77%. Huawei specifically emphasized that this system solves the pain point of traditional radars where small objects are detected but not trusted at high speeds of 120 km/h. In addition, the LiDAR adopts an industry-first tempered film glass window, which increases the window hardness by 25% and durability by 2 times, solving the problem of easy wear and tear of the LiDAR window in actual use to a certain extent.
In fact, from the technological routes of Hesai Technology and Huawei, we can see that Hesai is trying to enable the LiDAR to see colors by itself, thereby replacing some of the post-fusion work; while Huawei uses high resolution and a dual-focus architecture to make the point cloud itself sufficiently fine, allowing the algorithm to directly extract features from it without relying on massive post-processing.
Chips Are Becoming the Core of Competition
In the past, when people talked about LiDAR, they focused on whether the mechanical structure was rotating or solid-state, or whether the wavelength was 905 nm or 1550 nm. But by 2026, chip integration has become a more focused topic. The industry's view on the trend of LiDAR is also converging: LiDAR is undergoing a transition from analog architecture to digital architecture, and future competition will revolve around chip capabilities.
RoboSense is one of the earliest companies to productize the concept of digitalization. Its EM4 is the world's first mass-produced thousand-line ultra-long-range digital product, equipped with SPAD-SoC and VCSEL chips, integrating technologies such as crosstalk elimination, full-condition optoelectronic signal processing, and lossless data compression. Based on platform-based design, the EM4 can be precisely customized with different specifications such as 520 lines, 720 lines, 1080 lines, and 2160 lines according to the needs of automakers.
Taking the 1080-line version as an example, its angular resolution reaches 0.050°×0.025°, with a maximum detection range of 600 meters and a point frequency of 25.92 million points/second. In real-world tests, it can clearly detect targets such as tires at 180 meters, black cardboard boxes at 250 meters, and traffic cones within 300 meters. Compared with current mainstream LiDAR products, the EM4 can increase the system response time by up to 70%. Paired with the E1 full solid-state blind-spot filling LiDAR (FOV 120°×90°), this combination can clearly identify small targets of 13x17 cm within a distance of 130 meters, covering near-field low obstacle detection. RoboSense's E1 full solid-state blind-spot filling LiDAR is currently the only mass-producible automotive-grade full solid-state blind-spot filling LiDAR in the industry. The two work together to form an omnidirectional perception solution from long-range precise recognition to near-range blind-spot elimination.
At its 2026 Tech Day, RoboSense further clarified its digitalization direction and released a digital architecture named Genesis. This architecture elevates the output of LiDAR from sparse point clouds to imaging-grade levels, enabling the radar itself to have stronger information output capabilities. The combination of EM4 and E1 has been used by over 90% of the world's leading Robotaxi companies and has been integrated into the NVIDIA DRIVE AGX computing platform, making it directly compatible with NVIDIA's autonomous driving development system.
Hesai also has profound accumulation in the chip field. On November 24, 2025, Hesai released the Fermi C500, a high-performance intelligent main control chip dedicated to LiDAR based on the RISC-V architecture. The Fermi C500 adopts an independent and controllable RISC-V architecture, integrating an MCU (Microcontroller Unit), FPGA, and ADC on a single chip. It is the world's first LiDAR-dedicated main control chip that integrates dual certifications for functional safety and cybersecurity on a single chip. Functional safety certification is extremely critical in autonomous driving systems, ensuring that the chip can operate according to predetermined safety mechanisms in the event of a failure, without causing the collapse of the entire perception system.
Image source: Hesai Technology
The Fermi C500 also has a built-in Point Cloud Intelligent Processing Engine (IPE), integrating 256-core waveform processing cores. It can intelligently filter out environmental noise, allowing the LiDAR to still output high-precision perception data under complex weather conditions such as rain, fog, and dust. Hesai simultaneously released the Photon Isolation Security Technology, which suppresses the broadening artifacts of point clouds at the physical level, achieving precise separation of real signals and noise. This technology has been fully deployed in Hesai's main LiDARs such as ATX and ETX. The refreshed version of ATX, equipped with the Fermi C500 chip and Photon Isolation Technology, is a 256-line long-range LiDAR designed to meet the safety requirements of the combined driving assistance system for L2-level intelligent connected vehicles.
As of November 2025, Hesai has obtained AEC-Q automotive-grade certification for 16 self-developed chips, with a cumulative delivery of 185 million self-developed chips and semiconductor devices, ranking first globally. Hesai has completed the full-stack self-development of seven key components: lasers, detectors, laser drivers, TIA chips, ADC chips, digital signal processors, and controllers.
For LiDAR, the impact brought by chip integration is actually profound. On the one hand, it changes the cost structure of LiDAR, where expensive optical components and discrete components of the past are replaced by integrated semiconductor solutions; on the other hand, it amplifies the technological generation gap between different enterprises. Companies capable of self-developing core chips will gain advantages in system iteration speed and customization capabilities.
What New Choices Are There for Technical Architecture and Scanning Methods?
There have always been multiple choices for the technological route of LiDAR. In terms of wavelength selection, 905 nm and 1550 nm, as the two mainstreams, each have their pros and cons. 905 nm has a lower cost but slightly inferior eye safety, while 1550 nm offers a longer detection distance and stronger anti-interference capability but at a higher price. In terms of scanning methods, solutions such as mechanical, hybrid solid-state, and full solid-state coexist, with different trade-offs among reliability, cost, and performance for each solution.
Seyond's Falcon K3 LiDAR adopts the 1550 nm wavelength. Based on the technical foundation of self-developed core components, it has achieved performance breakthroughs, upgrading the maximum detection distance to 900 meters, and the standard ranging at 10% reflectivity to 450 meters. The angular resolution has improved by over 300% compared to the previous generation Falcon K2, with the ROI angular resolution reaching up to 0.05°×0.04°, capable of perceiving and detecting small targets of 20x30 cm at 250 meters away.
LiDAR with a 1550 nm wavelength inherently possesses excellent anti-interference capabilities and eye safety characteristics. Relying on its outstanding dynamic range and this feature, the Falcon K3 can accurately identify small objects 15 cm high at a distance of 160 meters. At the same time, based on the technical foundation of self-developed core components, the Falcon K3 has further improved its integration, reducing its volume by 50% compared to the previous generation, making it easier to adapt to integrated vehicle body designs.
Seyond has built three major platforms: Falcon, Robin, and Hummingbird. Falcon focuses on the 1550 nm high-performance route; the Robin series achieves high cost-performance configurations with the 905 nm solution (among which the Robin E2 can support up to 3400-line customized solutions); and the Hummingbird is a full solid-state near-field blind-spot filling product.
Because the full solid-state solution has no mechanical rotating parts, it theoretically has higher reliability and a longer service life. RoboSense's E1 belongs to the full solid-state blind-spot filling LiDAR, possessing an ultra-wide FOV of 120°×90°, specifically responsible for eliminating near-field blind spots. The EM4 is responsible for long-range high-precision detection, and the E1 is responsible for near-field large-range blind-spot filling. The two work together to form a 360° omnidirectional blind-spot-free perception solution. Relying on the mass production and application of the E1, RoboSense has accumulated significant advantages in this direction. The E1 adopts a self-developed SPAD-SoC chip, achieving the integrated integration of optoelectronics and computing. The entire transmitting and receiving system has no moving parts, further reducing the failure rate.
Huawei's 896-line LiDAR belongs to the hybrid solid-state route in terms of scanning method, but its dual-optical-path architecture itself is a more original design. This architecture solves the contradiction of traditional radars having to choose between seeing far and seeing wide. Through the coordinated work of two independent receiving units, it simultaneously obtains long-distance details and near-field wide-field information. The two units are each equipped with an independent SPAD chip, which can be controlled independently and work collaboratively, making the density of the point cloud output by the radar leading in the industry.
Final Words
From the technological layouts of these companies, it can be seen that current LiDAR is no longer purely pursuing long-range vision, but rather accurate and comprehensible perception. Chip integration brings the cost and performance iteration of the whole machine closer to the pace of the semiconductor industry. Multi-dimensional perception (color, ultra-high line count) enables the point cloud itself to carry more semantic information. The coexistence of different architectures such as 1550 nm and 905 nm, hybrid solid-state and full solid-state, indicates that the industry has not yet converged to a single technological route. However, what is certain is that LiDAR has moved from the debate stage of whether to install it to the stage of how to make it more usable.