Skip to content

Zoox Autonomous Vehicle Technology Stack — Deep Dive

Compiled March 2026 from zoox.com, NHTSA filings, California DMV reports, AWS re:Invent talks, Amazon Science publications, NVIDIA blog posts, patent filings, job postings, and trade press.


Table of Contents

  1. Company Overview
  2. Vehicle Platform
  3. Sensor Suite
  4. Onboard Compute
  5. Autonomy Software Stack
  6. Machine Learning & AI
  7. Mapping & Localization
  8. Simulation Platform
  9. Cloud & Data Infrastructure
  10. Programming Languages & Tools
  11. Safety & Redundancy Architecture
  12. Manufacturing & End-of-Line Testing
  13. Fleet Operations & Teleoperation
  14. Regulatory & Safety Record
  15. Key Partnerships & Suppliers
  16. Competitive Comparison
  17. Research & Publications
  18. Engineering Organization
  19. Sources

Company Overview

AttributeDetail
Founded2014
FoundersJesse Levinson (CTO), Tim Kentley-Klay
CEOAicha Evans (joined 2019; formerly Chief Strategy Officer at Intel)
CTOJesse Levinson — Ph.D. Stanford CS; algorithms for the $1M-winning 2007 DARPA Urban Challenge entry
Acquired by AmazonJune 2020, ~$1.2B + $100M retention
Pre-acquisition funding$990M across 6 rounds; last valuation $3.2B (2018)
Employees~2,300 (late 2025)
HQ1149 Chess Drive, Foster City, CA 94404
Other officesSan Francisco, Las Vegas, Boston (Strio.AI acquisition 2022), San Diego
ManufacturingHayward, CA (220,000 sq ft, opened June 2025); Fremont, CA (test fleet)

Key Milestones

  • 2015 — First autonomous trip
  • 2016 — Four-wheel steering validated
  • 2017 — Urban autonomous driving in San Francisco; NVIDIA partnership begins
  • 2019 — Testing begins in Las Vegas; first crash test iteration
  • 2020 — CA DMV driverless testing permit (4th company ever); Amazon acquisition; vehicle revealed
  • 2021 — Testing expands to Seattle
  • 2022 — Strio.AI acquired (Boston robotics/AI)
  • 2023 — First fully autonomous public road journey with passengers; NeurIPS paper (Scenario Diffusion)
  • 2024 — Deployed on SF streets and Las Vegas Strip
  • 2025 Sep — Las Vegas public robotaxi service launches (free rides)
  • 2025 Nov — San Francisco public launch ("Zoox Explorers" program)
  • 2025 Aug — First-ever NHTSA demonstration exemption for American-built AVs

Vehicle Platform

Zoox is the only robotaxi company operating a ground-up purpose-built vehicle — not a retrofit.

SpecificationValue
ClassificationPassenger car (49 CFR Part 571.3)
DesignBidirectional, symmetrical — no fixed front or rear
SAE LevelDesigned for Level 5; operating as Level 4 (geo-fenced)
Length3,630 mm / 142.9 in (~12 ft)
Width~1,830 mm (~6 ft)
Height1,936 mm / 76.2 in (~6.3 ft)
Curb weight~5,400 lbs (2,449 kg)
Top speed75 mph (121 km/h) — achievable in either direction
Turning circle8.6 m (28.2 ft)
SteeringZF four-wheel steering — bidirectional, tight-space maneuvering
BrakingZF Integrated Brake Control (IBC) — electro-hydraulic, redundant fallback
Seating4 passengers, face-to-face carriage-style
ControlsNone — no steering wheel, no pedals, no mirrors
Battery133 kWh — two independent packs (one under each seat row)
Battery cellsPanasonic Energy 2170-format cylindrical Li-ion (multi-year deal, starting early 2026)
RangeUp to 16 continuous hours on a single charge
DrivetrainDual electric motors, all-wheel drive, fully redundant
Shape"Squircle" — rounded, symmetrical

Interior

  • Touchscreen at every seat
  • Wireless charging at every seat
  • Individual climate controls per seat
  • Spatial audio system
  • Ambient cabin lighting
  • Automatic carriage-style sliding doors
  • Moonroof / skylight
  • Two-way audio with Rider Support

Sensor Suite

Five sensor modalities in identical pods at each of the four vehicle corners, providing overlapping 360-degree coverage. Each corner achieves 270-degree FOV.

SensorSupplierKey SpecsRole
LiDARHesai TechnologyMultiple units per vehicle; likely AT128 (128-ch, 200m range, 1.53M pts/sec)3D geometry, precise distance measurement
CamerasNot disclosed~28 RGB cameras; wide and telephoto lensesColor, traffic lights, pedestrian gestures, classification
LWIR / ThermalTeledyne FLIRBoson modules (640×512, 12μm uncooled, 30/60 Hz); Intel Movidius Myriad 2 VPUHeat signatures — people/animals, day and night
RadarNot disclosedUnits at each cornerVelocity, long-range, adverse weather, penetrates occlusions
MicrophonesNot disclosedDirectional audioEmergency vehicle sirens, approach direction

Additional sensors for localization: GPS, accelerometers, gyroscopes, wheel speed sensors, steering angle sensors.

Total sensor count: ~64 sensors across the vehicle (described as "dozens").

Detection range: 150–200+ meters in all directions.

Sensor fusion approach: Early fusion — raw data from all modalities combined before independent object detection (not late fusion of separate detections). Temporal information incorporated for velocity estimation and scene flow.

Sensor Staleness Innovation (deployed summer 2025)

Zoox developed a model-agnostic framework adding timestamp features to every data point for fine-grained temporal awareness. Trained using synthetic stale data from real-world logs.

  • Pedestrian detection precision doubled
  • Recall increased ~600%
  • Near-zero latency impact

Onboard Compute

ComponentDetail
GPUsMultiple NVIDIA GPUs on the NVIDIA DRIVE platform
CPUs4× Intel Xeon processors
Original platformNVIDIA DRIVE PX Pegasus (320+ TOPS)
Current platformLikely DRIVE AGX Orin or Thor generation (exact model undisclosed)
ArchitectureCentralized — all perception, prediction, planning, and control on one compute platform
RedundancyDual mirrored computer systems in vehicle floor with cross-verified logic domains
Data generationUp to 4 TB/hour of raw sensor data per vehicle
OTA updatesContinuous over-the-air neural network and software updates (typically every few weeks)

CTO Jesse Levinson: "We've been using NVIDIA hardware since the very start... a couple of orders more magnitude of computation done with the same amount of power" over the past decade.


Autonomy Software Stack

Pipeline: Perception → Localization → Prediction → Planning → Control

Perception (Three Independent Parallel Systems)

SystemTypeDescription
Main AI SystemML-basedDetection, classification, tracking, segmentation across all 5 sensor modalities
Geometric Collision AvoidanceNon-ML / interpretableDirect path-obstruction detection using geometric algorithms; 360-degree, low latency
"Safety Net"ML-basedIndependent collision-avoidance with short-horizon prediction; triggers emergency stop when collision probability exceeds thresholds

Prediction

AttributeDetail
Legacy systemUAP (Unified Active Prediction) — graph-based neural network
Current systemQTP (Query-centric Trajectory Prediction) — data-driven behavior modeling
Neural networksCNNs for bird's-eye-view scene (~60 semantic channels); GNNs for agent interaction via message passing
Prediction horizonUp to 8 seconds into the future
Update rateRecalculated every 100 milliseconds
Training dataBillions of real-world samples; self-supervised (actual future trajectories as ground truth)
Conditional predictionPredicts how other agents respond to Zoox's planned actions

Planning

AttributeDetail
FrameworkCost-based multi-objective optimization
ObjectivesSafety, rules of the road, journey completion, efficiency, rider comfort
ApproachHybrid — traditional motion planning + ML-based trajectory generation
Hard constraintsOverride all cost calculations (cannot leave drivable surface, wrong-way, etc.)
Update frequencyMultiple times per second
Training inputsProfessional driver "ideal driving" datasets + simulation-refined policies
BackupIndependent Collision Avoidance System (CAS) — millisecond-level response, assumes full control if primary planner fails

Machine Learning & AI

Foundation Model (presented AWS re:Invent 2025)

AttributeDetail
ArchitectureMultimodal language-action model with LLM core
Base modelQwen 2/3 VL (vision-language model)
InputsCamera/video (pre-trained vision encoders), LiDAR, radar, text prompts, existing perception outputs
OutputsRobotic controls (acceleration, braking, steering), 3D detections, generative responses
Model sizes400M → 7B → 32B parameters
Training Stage 1Large-scale supervised fine-tuning (behavior cloning on tens of thousands of hours of human driving)
Training Stage 2High-quality SFT (rare objects, difficult scenarios, synthetic chain-of-thought)
Training Stage 3Reinforcement learning using GRPO and DAPO techniques

ML Frameworks & Training

ToolUse
PyTorchPrimary deep learning framework
TensorFlow / KerasSecondary ML framework
JAXResearch / experimentation
NeMo / MegatronLarge model training
RayScalable distributed AI computation
DeepSpeedDistributed training optimization
Comet MLExperiment tracking
Mosaic Data Streaming (MDS)Deterministic, resumable data loading with mid-epoch resumption

Distributed Training Configuration

  • HSDP (Hybrid Sharded Data Parallel) + DDP across nodes
  • FSDP (Fully Sharded Data Parallel) within nodes
  • Tensor parallelism for large models
  • BF16 precision with gradient checkpointing
  • torch.compile for graph optimization
  • 64+ GPUs per training run; 500+ node clusters supported
  • 95% GPU utilization achieved after optimization

Model Inference

ToolContext
vLLMOffline model serving
TensorRT LLMOnline / on-vehicle inference (in development)

Key Model Architectures

  • CNNs — Bird's-eye-view perception (~60 semantic channels)
  • GNNs — Agent interaction modeling via message passing
  • Latent Diffusion Models — "Scenario Diffusion" for synthetic scenario generation (NeurIPS 2023)
  • Gaussian Splatting / NeRFs — Neural rendering for 3D scene reconstruction in simulation

Annotation & Labeling

  • Dedicated Perception Labeling & Tools team
  • Auto-labeling algorithms reduce manual burden
  • Self-supervised prediction training (actual future trajectories as labels)
  • Web-based labeling tools (React/Angular/Vue + Python backends)

Mapping & Localization

CLAMS (Calibration, Localization, and Mapping Simultaneously)

  • Drives Toyota Highlander survey vehicles through target areas
  • ML-based dynamic object removal (people, vehicles, temporary objects)
  • Produces HD 3D point-cloud maps from overlapping sensor data
  • Infrastructure-free calibration — uses natural environmental features (building edges, tree trunks aligned between camera gradients and LiDAR depth edges)

ZRN (Zoox Road Network)

  • Semantic layer atop HD 3D maps
  • Encodes: speed limits, traffic signals, stop signs, bike lanes, one-way streets, keep-clear zones
  • ZRN Monitor — real-time detection of discrepancies between map and real world (construction, moved lane markings); alerts fleet and engineering

Localization

AttributeDetail
Position accuracyWithin a few centimeters
Heading accuracyWithin a fraction of a degree
Update rate200 times per second
Sensors usedLiDAR, cameras, GPS, accelerometers, gyroscopes, wheel speed, steering angle
MethodMatches real-time sensor data against HD maps

Simulation Platform

AttributeDetail
EngineCustom-built C++ simulator
ScaleMillions of scenarios daily
Scenario typesEngineered (human-designed), log-based (real-world replay), system-generated (procedural)
Generative modelsScenario Diffusion (latent diffusion) — generates synthetic driving scenarios from noise in ~1 sec/scenario on a single GPU
Adversarial testingAutomated adversarial simulations for safety-critical edge cases
World creationProcedural environments with Houdini
Neural renderingGaussian Splatting, NeRFs for 3D reconstruction
ComputeLarge GPU clusters; can reserve 2,000 GPUs via AWS EC2 Capacity Blocks

Physical Testing — Altamont Test Track

  • Main track — high-speed testing
  • Inner loop — city intersection simulation
  • Test pad — lateral movement testing

Hardware-in-the-Loop (LabBot)

  • DynoBot — full driving components + dynamometers simulating real-world movement
  • GoldenBot — stationary electrical architecture integration testing
  • Fault injection (software bugs, brake failures), millisecond-precision logging

Cloud & Data Infrastructure

Hybrid Architecture: On-Premises + AWS

On-Premises

ComponentDetail
GPU clusterThousands of NVIDIA GPUs (supercomputer class)
StorageQuobyte parallel filesystem — 3 clusters, 30 PB, tens of thousands of concurrent clients (migrated from Ceph)
TieringSSD → disk drives → cloud
Training cadence~Every two weeks

AWS Services

ServiceUse
EC2 (P5, P6N GPU instances)ML training compute
EC2 Capacity BlocksReserve up to 2,000 GPUs for simulation/training
SageMaker HyperPodDistributed training with auto-recovery and health checks
EKSKubernetes orchestration (thousands of instances)
S3Primary object storage — tens of PB active, ~1 exabyte cold
FSx for LustreHigh-performance parallel filesystem for training
EFSShared filesystem
CloudWatchMonitoring
Managed Grafana + PrometheusObservability stack
EFA (Elastic Fabric Adapter)3,200 Gbps/node inter-node networking
AWS Data Transfer TerminalsPhysical upload at up to 400 Gbps from vehicles

Orchestration

  • Slurm (SchedMD) — current workload manager
  • Transitioning to EKS-based SageMaker HyperPod
  • Can spin up 1,000 nodes within a single AWS Region for burst

Scale

  • ~1 exabyte of total data (cold + active)
  • 4 TB/hour generated per vehicle
  • 500+ node training clusters
  • 95% GPU utilization achieved

Programming Languages & Tools

Languages

LanguageUsage
C++ (modern)Core autonomy stack, middleware, simulator, real-time embedded, safety-critical systems
PythonML training, data pipelines, backend services (FastAPI, Django, Flask), scripting
TypeScript / JavaScriptWeb tools, operational dashboards, 3D visualization (React, Vue.js)
JavaMentioned in interview requirements
SQLData querying across platforms

Build & Dev Tools

ToolUse
BazelPrimary build system
GitVersion control
DockerContainerization
TerraformInfrastructure as code
Kubernetes / EKSOrchestration

Data Engineering

ToolUse
Apache Spark / DatabricksLarge-scale data processing
Apache AirflowWorkflow orchestration
Kafka / KinesisStreaming data
RayDistributed compute

Frontend / Visualization

ToolUse
React (primary), Vue.js, AngularWeb frontends for internal tools
three.js / Babylon.js3D rendering
Vulkan / OpenGL3D graphics APIs

Monitoring & Observability

ToolUse
Prometheus + GrafanaMetrics and dashboards
NVIDIA DCGM ExporterGPU-level monitoring
CloudWatchAWS monitoring

RTOS & Embedded

  • FreeRTOS, SafeRTOS, QNX, Linux

Systems Engineering

  • DOORS, JAMA, Polarion (requirements management)
  • SysML (system modeling)

Middleware

Zoox develops proprietary custom middleware (not ROS in production):

  • Robot state machine
  • Software and message interfaces
  • Task schedulers and data transport layers
  • Diagnostic reporting
  • On-vehicle C++ in a real-time Unix-like environment

Safety & Redundancy Architecture

Design Philosophy: Fail-Operational (not merely Fail-Safe)

Inspired by aviation safety standards (ARP4754A, ARP4761). The vehicle continues operating safely through faults, rather than simply shutting down.

Safety standards referenced: FMVSS, ISO 26262, ISO 21448 (SOTIF), ISO 12207, DO-178, SPICE/ASPICE.

Redundancy Matrix

SystemRedundancy Approach
SteeringDual steering platforms (primary + backup) for bidirectional control
BrakingMultiple backup functions, multi-technology approach, third independent emergency brake
PowerTwo independent battery packs + redundant power converters + two additional 12V backup batteries
ComputeDual mirrored computer systems in vehicle floor with cross-verified logic domains
Sensors360° overlapping FOV; operational if individual sensors fail
ConnectivityThree cellular modems operating simultaneously with load balancing
AutonomyMain driving stack + independent Collision Avoidance System
DrivetrainTwo independent electric motors

Post-failure capability: Vehicle can pull over safely, activate hazard lights, open doors, and shut down HVAC even with battery or converter failure.

Crash Safety

  • 100+ safety innovations not found in conventional vehicles
  • Horseshoe (U-shaped) airbag system — wraps 180° around passengers; protects from front and side (industry first for carriage seating)
  • Five distinct airbag types: horseshoe curtain, frontal (split head/neck/chest), rear, side head, seat side
  • Intelligent deployment — control unit detects collision direction and severity; deploys only relevant airbags in sequence
  • Novel crumple zones — driving module and motor assembly dissipate energy before reaching passenger carriage
  • Smart seatbelts with monitoring
  • Safety-focused active suspension — continuously adapts to road conditions
  • Target: five-star equivalent crash safety for every occupant
  • CAE simulation run thousands of times before physical prototyping (crash duration ~300 ms)

Manufacturing & End-of-Line Testing

Hayward Production Facility (opened June 2025)

AttributeDetail
Size220,000 sq ft
Capacity10,000+ robotaxis/year at full scale
Current rate~1 vehicle/day; target 3/hour
Workforce~100 technicians (growing to hundreds)
Assembly time~20 minutes per vehicle
ApproachModular — major components pre-assembled by suppliers; Zoox does final integration
AutomationRobots handle adhesive dispensing, AGVs transport vehicles; humans do most assembly
NotableNo welding, cutting, or painting on-site — low power draw vs. traditional auto plants

End-of-Line Testing Sequence

  1. Sensor calibration bay — automated turntable, halogen lights, radar targets; calibrates LiDAR, cameras, IR, radar
  2. Wheel & headlight alignment — both ends (bidirectional); active suspension calibration; electronic steering zeroed
  3. Dynamometer testing — autonomous stress tests up to 75 mph; laser-based lateral drift correction
  4. Water leak / rain simulation — validates seals and sensor cleaning system (water spray + air blasts)
  5. Light tunnel & appearance — exterior lights, doors, speakers, touchscreens
  6. Factory Static Test — VIN verification, safety faults, seatbelts, emergency release, two-way audio, NHTSA label
  7. Buzz/Squeak/Rattle on outdoor test track
  8. Factory Dynamic Test — extended autonomous closed-loop driving (hours), both clockwise and counterclockwise

Fleet Operations & Teleoperation

Three Operations Teams (Foster City HQ)

TeamRole
Mission Operations"Air traffic controllers" — fleet health (tire pressure, battery, cabin/coolant temps), rerouting around street closures, demand-based fleet reallocation, weather advisories
TeleGuidanceNot remote driving — when vehicle encounters unfamiliar scenario, tactician draws waypoint "breadcrumbs" on screen; vehicle follows path autonomously while maintaining its own safety responsibility; response within seconds
Rider SupportFull ride lifecycle — start/end checks, seatbelt verification, lost belongings, in-app messaging, emergency button response

Ride-Hailing App

  • iOS and Android
  • Request ride → estimated wait time and trip duration
  • Physical "Zoox concierge" staff at pickup points
  • In-vehicle touchscreen for trip progress, music, temperature
  • One-tap live support
  • Currently free; paid rides planned Las Vegas early 2026, SF later 2026

Operational Cities

StatusCities
Public serviceLas Vegas (Sep 2025), San Francisco (Nov 2025)
TestingAustin, Miami, Seattle, Atlanta, Los Angeles, Washington D.C., Columbus OH
Next launchesAustin, Miami

Fleet size: ~50 purpose-built robotaxis + hundreds of Toyota Highlander test vehicles.


Regulatory & Safety Record

NHTSA Exemptions

ExemptionDetail
Demonstration Exemption (Aug 4, 2025)First-ever for American-built AVs under expanded AVEP; up to 2,500 vehicles/year
Temporary Exemption (filed Jun 2025)Pending; requests exemption from FMVSS 103, 104, 108, 111, 135, 201, 205, 208 (human-driving aids inapplicable to driverless bidirectional vehicle)

NHTSA Recalls

RecallDateVehiclesIssueRemedy
25E-019Mar 2025258Over-cautious hard braking for bicyclists near crosswalks; incorrect collision anticipation from rearOTA update to v24.32
25E-037May 2025270At <0.5 m/s, failure to detect prone VRU immediately adjacentOTA update (deployed May 21, 2025)
25E-090Dec 2025332Unnecessary lane-line crossing at/near intersections (62 instances, zero collisions)Two-phase OTA update

Fleet growth from recalls: 258 → 270 → 332 vehicles over 2025.

California DMV Testing Data

PeriodRegistered VehiclesMiles (safety driver)Driverless MilesDisengagementsMiles/Disengagement
2023 (Dec 2022 – Nov 2023)281710,40911,2634~177,602
2024 (Dec 2023 – Nov 2024)380951,87137,804

Safety Record

  • 100+ million fully autonomous miles accumulated
  • As of Jan 2026: 116 NHTSA-logged incidents in autonomous mode (small number resulted in injury/property damage)
  • May 2024: NHTSA investigation after two rear-end collisions with motorcycles

Key Partnerships & Suppliers

PartnerRole
AmazonParent company (acquired 2020, ~$1.3B); AWS infrastructure
NVIDIADRIVE platform GPUs (on-vehicle + data center); partnership since 2017
Hesai TechnologyLiDAR sensors (multi-year partnership)
Teledyne FLIRLongwave infrared (thermal) cameras
Panasonic Energy2170-format Li-ion battery cells (multi-year deal, starting early 2026)
ZFFour-wheel steering, chassis modules, occupant safety/protection systems
IntelXeon CPUs (4 per vehicle); Movidius VPU in FLIR modules
MicrochipAutomotive chips
Formula One Williams RacingMulti-season partnership (simulation expertise sharing)
Las Vegas Golden KnightsMulti-year collaboration
Strio.AIAcquired 2022 for robotics/AI automation (Boston R&D)

Competitive Comparison

DimensionZooxWaymoTeslaCruise
VehiclePurpose-built, no steering wheelRetrofitted Jaguar I-PaceModified Model Y (CyberCab planned)Retrofitted Chevy Bolt EV (paused)
SensorsLiDAR + radar + cameras + LWIR + mics (~64)13 cameras, 4 LiDARs, 6 radars, audioCamera-only (Tesla Vision)LiDAR + radar + cameras
Unique sensorLWIR thermal cameras (first in industry)
Autonomy levelL4 geo-fencedL4 geo-fencedL4 geo-fenced (Austin pilot)L4 geo-fenced (paused)
Operational cities2 (LV, SF)4+ (Phoenix, LA, SF, Austin)1 (Austin pilot)Paused after 2023 incident
Fleet size~50 robotaxis + hundreds of test vehicles700+Consumer vehiclesGM-owned fleet (paused)
PricingFree (paid 2026)~$20/trip~$4/ride (Austin)N/A
BackingAmazonAlphabetTeslaGM (scaled back)
PerceptionTriple-redundant (AI + geometric + safety net)Multi-modal fusionNeural net vision-onlyMulti-modal fusion
Vehicle designFull vertical integrationRetrofit + custom sensor podMass-market vehicle + softwareRetrofit

Key Zoox Differentiators

  1. Only company with a purpose-built vehicle (no retrofit compromises)
  2. Only company using LWIR thermal cameras in the sensor stack
  3. Triple-redundant perception (AI + geometric + safety net)
  4. Bidirectional driving eliminates three-point turns
  5. Full vertical integration from vehicle design through fleet operations
  6. Amazon/AWS infrastructure advantage for simulation and data processing
  7. Fail-operational (not just fail-safe) — aviation-inspired redundancy

Research & Publications

PublicationVenueTopic
Scenario DiffusionNeurIPS 2023Controllable driving scenario generation with latent diffusion models; autoencoder + diffusion architecture; ~1 sec/scenario on single GPU
Sensor Staleness FrameworkAmazon Science 2025Model-agnostic temporal feature engineering; pedestrian precision 2×, recall 6× improvement
QTP (Query-centric Trajectory Prediction)Amazon ScienceData-driven behavior prediction replacing UAP
CLAMS LocalizationAmazon ScienceInfrastructure-free calibration using natural features
Foundation Model for AVAWS re:Invent 2025 (AMZ304)Multimodal language-action model (Qwen VL base) outputting robotic controls

Conference Appearances

  • AWS re:Invent 2025 — "Building Machine Learning Infrastructure for Autonomous Vehicles" (AMZ304)
  • NVIDIA GTC 2024 — CTO Jesse Levinson on NVIDIA partnership and simulation
  • Research published under the Amazon Science umbrella

Engineering Organization

Major Divisions

DivisionKey Leaders
Autonomy SoftwareVP: Marc Wimmershoff
PerceptionDirector: Bat-El Shlomo; also Ruijie He (Strio.AI)
Prediction
Planning & Control
SimulationScenario frameworks, 3D sensor sim, sim data platform, sim infra
ML InfrastructureCPU/GPU resources, HPC
Driving ToolsReal-time operational tools, 3D visualization
Developer Platforms
SDMA (Systems Design & Mission Assurance)V&V, safety metrics, safety clearance
FirmwareEmbedded systems
Manufacturing OperationsHayward and Fremont facilities
Safety Strategy & OperationsStandards compliance, safety cases
HomologationRegulatory filings
User ExperienceApp, in-vehicle experience

Locations: Foster City (HQ), San Francisco, Las Vegas, Boston (R&D), San Diego.


Sources

Official Zoox

Amazon Science

AWS & NVIDIA

Regulatory

Infrastructure

Trade Press


Compiled from 7 parallel research agents scanning 60+ sources across zoox.com, NHTSA, California DMV, Amazon Science, NVIDIA, AWS, patent databases, job postings, and trade publications.

Public research notes collected from public sources.