Sensors and Compute

cameras · radar · lidar · compute

Every autonomy stack is bounded by what its sensors can see and how fast its computer can decide. The cost of that hardware has collapsed, and China now makes most of the lidar.

Data path of an autonomous vehicle: cameras (eight to 13 per car), radar (front plus corners), lidar (zero to four units; Chinese suppliers Hesai and RoboSense lead volume) and inertial, GNSS and wheel sensors feed a network layer of camera serializer links (GMSL, MIPI A-PHY) and automotive Ethernet with time-sensitive networking. The network feeds a central compute module such as Nvidia DRIVE Thor, Tesla AI4 or AI5, or Horizon Robotics Journey 6, which commands braking and steering. Underneath sits a power and cooling layer with redundant rails, liquid-cooled compute and safe derating. SENSE, MOVE, DECIDE, ACT Cameras 8 to 13 per car Radar front plus corners Lidar zero to four units IMU, GNSS wheels, odometry Network GMSL, A-PHY Ethernet, TSN Compute Thor, AI4, AI5 Horizon J6 Act brake steer Power and cooling redundant rails, liquid-cooled compute, safe derating Red marks a part where Chinese suppliers lead volume (lidar: Hesai, RoboSense)
Cameras per car
8 to 13
Lidar, China
~$200 a unit
Default compute
Nvidia Thor
Hub
Autonomy

Finding

Autonomy hardware has split into two camps by sensor set, and both now run on a single central computer

Tesla and most Chinese urban driver-assistance systems rely on cameras with radar or no radar; Waymo, Baidu Apollo Go, Huawei, Rivian and Chinese premium brands add lidar. Either way, perception and planning have moved from many small ECUs onto one or two central AI computers such as Nvidia DRIVE Thor, Tesla AI4 or Horizon Robotics Journey 6.

Sensor cost is no longer the dividing line it was. Hesai shipped 1.62 million lidars in 2025 and sells its mass-market ATX unit for about $200, so Chinese carmakers fit lidar as standard on mid-priced cars. The open questions are compute and data: how many trillion operations per second fit inside a vehicle's power and cooling budget, how much camera video the network can carry without loss, and whether the redundancy design (overlapping sensors, dual compute paths, separate power rails) can support a Level 3 or Level 4 safety case under ISO 26262 and ISO 21448.

For the vehicle-level architecture see ADAS and AV Compute Architecture; for how fleets keep these sensors working see Fleet Autonomy Sensors and Perception.

Sensors

Sensor modalities

Autonomy sensor modalities
SensorStrengthLimitMain suppliersPage
CamerasColor, texture, signs, signals, lane linesGlare, darkness, dirt, no direct depthSony and OmniVision sensors; Sunny Optical and other Chinese module makersCameras
RadarRange and velocity in rain, fog, dust, nightLow angular resolution (improving with 4D imaging radar)Bosch, Continental, Aptiv, Arbe; Calterah and Huawei in ChinaRadar
LidarDirect 3D geometry, independent of lightCost (now about $200 in China), contamination, integrationHesai, RoboSense, Seyond (China); Valeo, Innoviz, OusterLiDAR
UltrasonicClose range in parking and dockingVery short rangeBosch, Valeo; dropped by Tesla and Rivian-
IMU, GNSS, encodersHigh-rate motion and absolute positionGNSS jamming and spoofing; driftBosch, STMicro, Murata; u-blox, Septentrio-
Force and torqueContact and balance in robotsCalibration, cost per jointATI, Robotiq; Chinese makers in humanoid jointsActuators

Compute

Autonomy compute platforms

TOPS figures are not comparable across vendors (precision and sparsity differ), so they are left out.

Autonomy compute platforms
PlatformDesignerUsed by
Nvidia DRIVE AGX ThorNvidiaMercedes-Benz, Volvo, BYD, Zeekr, Li Auto, Xiaomi; Lucid and Nuro robotaxi
Nvidia DRIVE OrinNvidiaPrevious generation; still in many Chinese and Western models
Tesla AI4 / AI5TeslaAll Tesla vehicles, Cybercab, Optimus; AI5 volume from 2027
Horizon Robotics Journey 6Horizon Robotics (China)Most Chinese brands, including BYD, Li Auto and Volkswagen China models
Huawei MDC (Ascend)Huawei (China)AITO, Luxeed, Stelato, Maextro and other HIMA brands
Mobileye EyeQ6MobileyeCamera-first ADAS at many Western automakers
Qualcomm Snapdragon RideQualcommBMW and others; ADAS plus cockpit on one chip

Architecture

Networks, power and redundancy

Autonomy hardware architecture
ElementCurrent practiceConstraint
Central computeOne or two AI computers replace many ECUs for perception and planningThermal density; single point of failure without redundancy
Camera linksSerializers (Analog Devices GMSL, MIPI A-PHY) carry raw video over coaxBandwidth per camera; EMI
BackboneMulti-gigabit automotive Ethernet with time-sensitive networking (IEEE 802.1 TSN)Deterministic latency for safety traffic
Power and coolingLiquid-cooled compute modules; derating when hotSustained load on the 12 V or 48 V system
RedundancyOverlapping sensors, dual compute paths, separate power rails, backup braking and steeringISO 26262 ASIL-D and L4 safety cases

Sources

Sources and scope

Scope
Sensors and Compute for Autonomy
Sources
Hesai 2025 results; Nvidia; Horizon Robotics; Tesla; ISO 26262 and ISO 21448; IEEE 802.1
Updated
October 2026