The Robotaxi industry is currently accelerating its transition from the technology validation and demonstration operation phase into a critical cycle of pre-installed mass production and large-scale commercialization. Autonomous driving hardware will determine the safety ceiling, performance boundaries, and deployment costs of the vehicles.
In current industry practice, the selection of sensor and main control chip solutions shows both divergence and consensus.
This article reviews the latest autonomous driving sensor configurations and computing platform solutions of nine Robotaxi players, including Apollo Go, Pony.ai, WeRide, Waymo, and Tesla (summarized in the figure below), and discusses the future development landscape and evolution direction.
The information is sourced from publicly available online materials and internal industry collections. If there are any omissions or errors, please feel free to point them out at any time.
01. Major Players
1. Waymo
Waymo's sixth-generation Robotaxi, Ojai, follows the design philosophy of "streamlined hardware and upgraded performance," with the total number of sensors reduced by 42% compared to the fifth generation.
The perception system includes 13 17-megapixel high-resolution cameras (presumed to be Sony IMX735), 4 self-developed LiDARs (including 1 roof-mounted main LiDAR with a 300-meter detection range and 3 short-range blind-spot LiDARs), and 6 self-developed imaging millimeter-wave radars (FCC-certified model RADE8C).
Additionally, it is equipped with EARs (External Audio Receivers) and a sensor cleaning system integrating heating, wiper, and water spray functions.
The computing platform is likely an X86 processor + self-developed accelerator + FPGA solution, and is currently evolving towards a more streamlined and reliable direction.
2. Apollo Go
The autonomous driving perception system of Apollo Go's sixth-generation RT6 includes 12 vision cameras (8 8-megapixel HD cameras + 4 3-megapixel surround-view cameras), 4 Hesai AT128 long-range LiDARs, initially equipped with 4 InnoLiDAR blind-spot LiDARs, and 6 Continental millimeter-wave radars (understood to be 2 ARS513 long-range radars + 4 SRR523 corner radars, which may be upgraded to sixth-generation products later), supplemented by 12 Horn Auto Electronics ultrasonic radars and 1 set of CH Navigation IMU+RTK combined navigation system.
The new generation models (expected to be the seventh generation) will be upgraded to a combination of "4 RoboSense EM4 1000-line long-range LiDARs + 6 RoboSense E1 fully solid-state blind-spot LiDARs."
The domain controller adopts dual redundant computing units, with a single unit computing power of 600 TOPS and a total computing power of 1200 TOPS, likely using an X86 architecture paired with an NVIDIA GTX 3080 Ti GPU.
3. Pony.ai
Pony.ai's seventh-generation Robotaxi adopts a multi-sensor fusion solution. The perception system covers 9 LiDARs (1 Seyond Falcon K2 long-range LiDAR, 4 Hesai AT series mid-range LiDARs, 4 RoboSense E1 blind-spot LiDARs), 14 cameras including wide-angle, telephoto, and infrared types (the front-view main camera is 8 megapixels, supplied by LCEye), and 4 millimeter-wave radars.
It is additionally equipped with 4 exterior microphones, 2 water wading sensors, 1 set of collision sensors, and a self-developed sensor cleaning device.
The domain controller is equipped with 4 NVIDIA Orin-X chips, with a total computing power of 1016 TOPS and a design life of 10 years/600,000 kilometers. Compared to the previous generation, the volume, weight, and power consumption are reduced by over 50%, and the cost is reduced by 80%.
4. WeRide
WeRide's GXR mass-produced Robotaxi is equipped with the iteratively upgraded Sensor Suite. The upgraded SS 8.0 is configured with 4 RoboSense EM4 1000-line long-range LiDARs and 3 RoboSense E1 blind-spot LiDARs, paired with a solution of about 12 cameras and 4D imaging millimeter-wave radars, supplemented by a GNSS+IMU combined navigation system and ultrasonic radars.
The domain controller adopts the Lenovo Automotive Computing HPC 3.0 platform, equipped with 2 NVIDIA DRIVE AGX Thor X chips, a total computing power of 2000 TOPS, 128GB LPDDR5X memory, and 900GB Micron automotive-grade solid-state storage.
5. DiDi
The R2 Robotaxi jointly created by DiDi and GAC Aion is equipped with 33 high-precision sensors. The perception layer includes 14 visible light cameras (6 8-megapixel mid-range, 1 8-megapixel long-range, 4 short-range blind-spot fisheye, 1 backup mid-range, 1 dedicated for traffic lights), 1 Raytron infrared camera, 4 long-range LiDARs (initially Hesai AT128P, later switched to RoboSense EM4), 6 RoboSense E1 blind-spot LiDARs, and 7 millimeter-wave radars (4 Saien Lingdong SFR-2K 4D main radars + 3 backup radars), and 2 acoustic sensors for identifying emergency vehicle siren sounds.
It is equipped with the full-stack self-developed "Orca" three-domain fusion computing platform, integrating the autonomous driving domain, cockpit domain, and communication & navigation domain. It is understood to adopt an X86 + 2 Orin chips architecture, with GPU computing power exceeding 2000 TOPS, CPU reaching 48 cores, and supporting active-standby redundant hardware architecture.
6. Hello Inc.
The HR1 Robotaxi, created by Zaofu Intelligence Technology under Hello Inc., is developed based on the Dongfeng Venucia VX6 redundant chassis. The perception system is equipped with 14 high-resolution cameras (including 4K HDR cameras), 4 Hesai AT128 mid-to-long-range LiDARs, 4 FTX blind-spot LiDARs, and 6 millimeter-wave radars, supplemented by 12 ultrasonic radars and collision sensors distributed in multiple places on the car body.
The domain controller adopts primary and secondary dual redundant computing units, equipped with dual NVIDIA Thor chips, with a total computing power exceeding 2200 TOPS, and plans to achieve SOP mass production in 2026.
7. XPeng
XPeng's GX Robotaxi adheres to the pure vision technology route, with no LiDAR and HD map configurations. It is presumed to follow the visual primary perception architecture of 11-12 external cameras, equipped with LOFIC Eagle Eye image enhancement technology to solve the imaging pain points of backlight and low-light scenarios from the hardware bottom layer, paired with millimeter-wave radars, ultrasonic radars, high-precision positioning units, and DMS/OMS systems.
The computing platform is equipped with 4 full-stack self-developed Turing AI chips, with a total computing power of 3000 TOPS. At the algorithm level, it adopts a dual-large-model architecture of the second-generation VLA driving large model + VLM cockpit large model.
8. Mobileye
The Mobileye Drive autonomous driving system adopts a "true redundancy" dual-independent subsystem architecture. The perception end uses 13 cameras (8 8-megapixel functional cameras, 1 backup camera, 4 2-megapixel surround-view cameras) to form a pure vision main link, paired with 3 Innoviz long-range LiDARs, 6 Innoviz Flash short-range LiDARs, and 5-6 self-developed software-defined 4D imaging millimeter-wave radars to form an independent redundant link, supplemented by GNSS+IMU and REM crowdsourced HD maps.
The computing platform is equipped with 4 EyeQ 6H SoCs, with a single chip computing power of about 67 TOPS. It supports independent computing of the two subsystems by domain, meeting high-level automotive functional safety requirements.
9. Tesla
Tesla's CyberCab adopts the pure vision technology route, with the perception core being 8 cameras. It is equipped with weather-proof lenses provided by Samsung (with built-in heating elements, coating hardness 6 times that of traditional cameras, and hydrophobic properties), paired with a dedicated wiper and water spray cleaning solution. Some versions retain 1 4D imaging radar, while some HW4.0 versions cancel the radar configuration, supplemented by ultrasonic radars and a GNSS+IMU positioning system.
The computing platform adopts the full-stack self-developed HW series solution. HW4.0 is equipped with the computing power of 2 self-developed AI4 chips, and HW5.0 is equipped with the computing power of the self-developed AI5 chip of about 2000-2500 TOPS, realizing autonomous driving capabilities in conjunction with the FSD end-to-end solution.
02. Hardware Development Trends
1. Cameras Upgrade to Ultra-High Pixels, 17 Megapixels Begin to Enter Vehicles
Traditional 8-megapixel automotive cameras have become the industry baseline, and leading manufacturers are beginning to break through to higher pixel specifications to improve long-distance recognition accuracy and low-light perception capabilities.
Waymo's sixth-generation Robotaxi (Ojai) is equipped with 13 high-resolution cameras across the vehicle, with the main sensor reaching up to 17 megapixels, which is the highest visual specification among currently mass-produced Robotaxis. Tensor (formerly the Robotaxi player AutoX), developing L4-level Robocars for individuals, has even adopted 17 17-megapixel cameras.
Compared with mainstream 8-megapixel cameras, 17-megapixel sensors possess higher resolution, wider dynamic range, and better low-light imaging performance. They can achieve longer detection distances and finer target recognition (such as long-distance traffic lights and small-sized obstacles) under the premise of controlling the number of cameras, and at the same time realize more accurate visual semantic fusion with LiDAR.
2. Continued Divergence in LiDAR Usage, 1000-Line Main LiDAR + Fully Solid-State Blind-Spot Becomes Mainstream Combination
Tesla and XPeng adhere to the solution without LiDAR and HD maps, relying on hardware-level vision enhancement + algorithm iteration to achieve L4 capabilities. More other players adhere to the multi-sensor fusion solution of LiDAR + vision + millimeter-wave radar, relying on hardware redundancy to ensure safety in extreme scenarios.
For example, Waymo's sensors can detect pedestrians behind buses and avoid collisions:
Long-range main LiDARs are generally upgraded to the 1000-line level, with the detection distance increased to 500-600 meters, and the point cloud density increased by several to more than ten times. The representative solution is RoboSense EM4, which has been designated or equipped on multiple new generation models such as Apollo Go's seventh generation, WeRide SS 8.0, and DiDi R2.
Blind-spot LiDARs are uniformly converging towards the Flash fully solid-state solution. RoboSense E1 and Hesai FTX have become the industry's general blind-spot choices, responsible for near-range blind-spot coverage of the car body, paired with long-range LiDARs to form 360° dead-zone-free perception. Multiple manufacturers are developing "color blind-spot LiDARs" that can simultaneously output RGB and depth information. Perhaps in the near future, surround-view cameras will be completely replaced.
Leading manufacturers are accelerating the cost reduction of self-developed LiDARs: Waymo and Huawei have both achieved full self-development of LiDARs, with deep customization at the optical and chip levels, balancing performance and cost.
3. Continued Iteration of Millimeter-Wave Radars, 4D Imaging and Satellite Architecture Become Important Directions for Perception Redundancy
4D imaging radars have significantly improved in angular resolution, point cloud density, and number of virtual channels compared to traditional radars, possessing height dimension perception capabilities. For example, Mobileye's self-developed imaging radar has over 1500 virtual channels, with an azimuth resolution of less than 0.5°; DiDi is equipped with Saien Lingdong SFR-2K 4D main radars and backup radars.
Some manufacturers, such as Mobileye, plan to gradually replace side and rear blind-spot LiDARs with 4D imaging radars, compressing hardware costs under the premise of ensuring safety redundancy.
The satellite radar solution can further leverage the advantages of radars. Some manufacturers have begun to explore the 16T16R satellite solution, which is expected to be applied to Robotaxis in the near future.
4. Perception Dimension Upgrade, From "Driving Perception" to "Full-Scenario Operational-Level Perception"
Facing the all-weather and long-cycle operation needs of Robotaxis, special sensors have become standard, such as the general addition of infrared night vision, environmental sound, water wading detection, and collision contact sensors. DiDi is equipped with infrared cameras and siren sound recognition sensors; Pony.ai is equipped with water wading sensors, elastic wave collision sensors, and exterior microphones; Waymo is equipped with EARs external audio receivers, which can identify emergency vehicle alerts and railway crossing warnings.
The sensor cleaning system is standardized to achieve all-weather autonomous cleaning and ensure normal working hours. Many LiDARs and cameras integrate cleaning devices such as heating, water spray, and dedicated wipers to cope with extreme weather such as rain, snow, sandstorms, and low temperatures, ensuring operational availability. Waymo, Pony.ai, and Tesla all have targeted cleaning and harsh weather resistance solutions.
5. Computing Platform Computing Power Enters the 2000 TOPS Level, Multi-Domain Fusion is Optional, Dual Redundancy is Standard
The total computing power of mainstream mass-produced Robotaxis has generally broken through 1000 TOPS, and high-end solutions reach 2000-3000 TOPS, supporting large model algorithms and massive data processing of multiple sensors. For example, XPeng is equipped with 4 self-developed Turing AI chips, with a total computing power of 3000 TOPS; Hello Inc. and WeRide adopt the dual Thor solution, with computing power reaching the 2000 TOPS level.
The architecture is evolving from a single autonomous driving domain controller to multi-domain fusion central computing. A typical example is DiDi's "Orca" platform, which realizes the three-domain fusion of the autonomous driving domain, cockpit domain, and communication & navigation domain.
Safety redundancy has become a hard requirement. Dual redundant computing units and active-standby dual architectures are the general design for all L4 solutions, ensuring that the vehicle can safely execute the minimum risk strategy when a single system fails.
There is still route divergence in main control chips. X86 and NVIDIA Orin/Thor are still common industry solutions; leading manufacturers are accelerating self-developed chips (XPeng Turing, Waymo self-developed SoC, Mobileye EyeQ, Tesla HW5 AI chips). With the advancement of automotive regulations, the X86 platform will gradually exit the historical stage.
6. Industry Landing Oriented, Transitioning from Modification Demonstration to Pre-Installed Mass Production, Costs Drop Significantly
The new generation of Robotaxis are all forward-developed based on mass-produced vehicle platforms, adopting pre-installed mass production solutions to replace the early post-installed modification mode. The reliability of the whole vehicle is significantly improved, adapting to large-scale operations.
The cost of the whole vehicle is accelerating its decline: Pony.ai has clearly stated that the total cost of the 2027 version of the L4 Robotaxi will be less than 230,000 RMB; Tesla's CyberCab target cost is controlled within 30,000 USD. Cost reduction has become the core proposition for commercial landing. Adopting automotive-grade components, reducing the number of components, and pre-installed solutions are key measures for cost reduction.
03. Conclusion
Overall, Robotaxi hardware solutions continue to evolve around the three core goals of "safety, performance, and cost." As high-level autonomous driving algorithms and end-to-end large models continue to penetrate, they support all-weather and full-scenario operational capabilities while continuously driving down the cost of the whole vehicle.
The synergistic evolution of future hardware capabilities and algorithm efficiency will become the core driving force for Robotaxis to achieve a wider range of daily urban travel.
The information in this article is sourced from publicly available online materials and internal industry collections. If there are any omissions or errors, please feel free to point them out at any time.
Personal views, not necessarily accurate, discussions are welcome.
I am Xueling, researching the technologies, products, and applications of perception, control, and AI. Welcome to exchange ideas.
To contact Xueling and for the full collection index, please refer to: https://dcn7get8fskg.feishu.cn/wiki/CCMpwjC0EiBIw2kFr7uc84qHneb