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
- Company Overview
- Vehicle Platform
- Sensor Suite
- Onboard Compute
- Autonomy Software Stack
- Machine Learning & AI
- Mapping & Localization
- Simulation Platform
- Cloud & Data Infrastructure
- Programming Languages & Tools
- Safety & Redundancy Architecture
- Manufacturing & End-of-Line Testing
- Fleet Operations & Teleoperation
- Regulatory & Safety Record
- Key Partnerships & Suppliers
- Competitive Comparison
- Research & Publications
- Engineering Organization
- Sources
Company Overview
| Attribute | Detail |
|---|---|
| Founded | 2014 |
| Founders | Jesse Levinson (CTO), Tim Kentley-Klay |
| CEO | Aicha Evans (joined 2019; formerly Chief Strategy Officer at Intel) |
| CTO | Jesse Levinson — Ph.D. Stanford CS; algorithms for the $1M-winning 2007 DARPA Urban Challenge entry |
| Acquired by Amazon | June 2020, ~$1.2B + $100M retention |
| Pre-acquisition funding | $990M across 6 rounds; last valuation $3.2B (2018) |
| Employees | ~2,300 (late 2025) |
| HQ | 1149 Chess Drive, Foster City, CA 94404 |
| Other offices | San Francisco, Las Vegas, Boston (Strio.AI acquisition 2022), San Diego |
| Manufacturing | Hayward, 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.
| Specification | Value |
|---|---|
| Classification | Passenger car (49 CFR Part 571.3) |
| Design | Bidirectional, symmetrical — no fixed front or rear |
| SAE Level | Designed for Level 5; operating as Level 4 (geo-fenced) |
| Length | 3,630 mm / 142.9 in (~12 ft) |
| Width | ~1,830 mm (~6 ft) |
| Height | 1,936 mm / 76.2 in (~6.3 ft) |
| Curb weight | ~5,400 lbs (2,449 kg) |
| Top speed | 75 mph (121 km/h) — achievable in either direction |
| Turning circle | 8.6 m (28.2 ft) |
| Steering | ZF four-wheel steering — bidirectional, tight-space maneuvering |
| Braking | ZF Integrated Brake Control (IBC) — electro-hydraulic, redundant fallback |
| Seating | 4 passengers, face-to-face carriage-style |
| Controls | None — no steering wheel, no pedals, no mirrors |
| Battery | 133 kWh — two independent packs (one under each seat row) |
| Battery cells | Panasonic Energy 2170-format cylindrical Li-ion (multi-year deal, starting early 2026) |
| Range | Up to 16 continuous hours on a single charge |
| Drivetrain | Dual 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.
| Sensor | Supplier | Key Specs | Role |
|---|---|---|---|
| LiDAR | Hesai Technology | Multiple units per vehicle; likely AT128 (128-ch, 200m range, 1.53M pts/sec) | 3D geometry, precise distance measurement |
| Cameras | Not disclosed | ~28 RGB cameras; wide and telephoto lenses | Color, traffic lights, pedestrian gestures, classification |
| LWIR / Thermal | Teledyne FLIR | Boson modules (640×512, 12μm uncooled, 30/60 Hz); Intel Movidius Myriad 2 VPU | Heat signatures — people/animals, day and night |
| Radar | Not disclosed | Units at each corner | Velocity, long-range, adverse weather, penetrates occlusions |
| Microphones | Not disclosed | Directional audio | Emergency 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
| Component | Detail |
|---|---|
| GPUs | Multiple NVIDIA GPUs on the NVIDIA DRIVE platform |
| CPUs | 4× Intel Xeon processors |
| Original platform | NVIDIA DRIVE PX Pegasus (320+ TOPS) |
| Current platform | Likely DRIVE AGX Orin or Thor generation (exact model undisclosed) |
| Architecture | Centralized — all perception, prediction, planning, and control on one compute platform |
| Redundancy | Dual mirrored computer systems in vehicle floor with cross-verified logic domains |
| Data generation | Up to 4 TB/hour of raw sensor data per vehicle |
| OTA updates | Continuous 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)
| System | Type | Description |
|---|---|---|
| Main AI System | ML-based | Detection, classification, tracking, segmentation across all 5 sensor modalities |
| Geometric Collision Avoidance | Non-ML / interpretable | Direct path-obstruction detection using geometric algorithms; 360-degree, low latency |
| "Safety Net" | ML-based | Independent collision-avoidance with short-horizon prediction; triggers emergency stop when collision probability exceeds thresholds |
Prediction
| Attribute | Detail |
|---|---|
| Legacy system | UAP (Unified Active Prediction) — graph-based neural network |
| Current system | QTP (Query-centric Trajectory Prediction) — data-driven behavior modeling |
| Neural networks | CNNs for bird's-eye-view scene (~60 semantic channels); GNNs for agent interaction via message passing |
| Prediction horizon | Up to 8 seconds into the future |
| Update rate | Recalculated every 100 milliseconds |
| Training data | Billions of real-world samples; self-supervised (actual future trajectories as ground truth) |
| Conditional prediction | Predicts how other agents respond to Zoox's planned actions |
Planning
| Attribute | Detail |
|---|---|
| Framework | Cost-based multi-objective optimization |
| Objectives | Safety, rules of the road, journey completion, efficiency, rider comfort |
| Approach | Hybrid — traditional motion planning + ML-based trajectory generation |
| Hard constraints | Override all cost calculations (cannot leave drivable surface, wrong-way, etc.) |
| Update frequency | Multiple times per second |
| Training inputs | Professional driver "ideal driving" datasets + simulation-refined policies |
| Backup | Independent Collision Avoidance System (CAS) — millisecond-level response, assumes full control if primary planner fails |
Machine Learning & AI
Foundation Model (presented AWS re:Invent 2025)
| Attribute | Detail |
|---|---|
| Architecture | Multimodal language-action model with LLM core |
| Base model | Qwen 2/3 VL (vision-language model) |
| Inputs | Camera/video (pre-trained vision encoders), LiDAR, radar, text prompts, existing perception outputs |
| Outputs | Robotic controls (acceleration, braking, steering), 3D detections, generative responses |
| Model sizes | 400M → 7B → 32B parameters |
| Training Stage 1 | Large-scale supervised fine-tuning (behavior cloning on tens of thousands of hours of human driving) |
| Training Stage 2 | High-quality SFT (rare objects, difficult scenarios, synthetic chain-of-thought) |
| Training Stage 3 | Reinforcement learning using GRPO and DAPO techniques |
ML Frameworks & Training
| Tool | Use |
|---|---|
| PyTorch | Primary deep learning framework |
| TensorFlow / Keras | Secondary ML framework |
| JAX | Research / experimentation |
| NeMo / Megatron | Large model training |
| Ray | Scalable distributed AI computation |
| DeepSpeed | Distributed training optimization |
| Comet ML | Experiment 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
| Tool | Context |
|---|---|
| vLLM | Offline model serving |
| TensorRT LLM | Online / 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
| Attribute | Detail |
|---|---|
| Position accuracy | Within a few centimeters |
| Heading accuracy | Within a fraction of a degree |
| Update rate | 200 times per second |
| Sensors used | LiDAR, cameras, GPS, accelerometers, gyroscopes, wheel speed, steering angle |
| Method | Matches real-time sensor data against HD maps |
Simulation Platform
| Attribute | Detail |
|---|---|
| Engine | Custom-built C++ simulator |
| Scale | Millions of scenarios daily |
| Scenario types | Engineered (human-designed), log-based (real-world replay), system-generated (procedural) |
| Generative models | Scenario Diffusion (latent diffusion) — generates synthetic driving scenarios from noise in ~1 sec/scenario on a single GPU |
| Adversarial testing | Automated adversarial simulations for safety-critical edge cases |
| World creation | Procedural environments with Houdini |
| Neural rendering | Gaussian Splatting, NeRFs for 3D reconstruction |
| Compute | Large 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
| Component | Detail |
|---|---|
| GPU cluster | Thousands of NVIDIA GPUs (supercomputer class) |
| Storage | Quobyte parallel filesystem — 3 clusters, 30 PB, tens of thousands of concurrent clients (migrated from Ceph) |
| Tiering | SSD → disk drives → cloud |
| Training cadence | ~Every two weeks |
AWS Services
| Service | Use |
|---|---|
| EC2 (P5, P6N GPU instances) | ML training compute |
| EC2 Capacity Blocks | Reserve up to 2,000 GPUs for simulation/training |
| SageMaker HyperPod | Distributed training with auto-recovery and health checks |
| EKS | Kubernetes orchestration (thousands of instances) |
| S3 | Primary object storage — tens of PB active, ~1 exabyte cold |
| FSx for Lustre | High-performance parallel filesystem for training |
| EFS | Shared filesystem |
| CloudWatch | Monitoring |
| Managed Grafana + Prometheus | Observability stack |
| EFA (Elastic Fabric Adapter) | 3,200 Gbps/node inter-node networking |
| AWS Data Transfer Terminals | Physical 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
| Language | Usage |
|---|---|
| C++ (modern) | Core autonomy stack, middleware, simulator, real-time embedded, safety-critical systems |
| Python | ML training, data pipelines, backend services (FastAPI, Django, Flask), scripting |
| TypeScript / JavaScript | Web tools, operational dashboards, 3D visualization (React, Vue.js) |
| Java | Mentioned in interview requirements |
| SQL | Data querying across platforms |
Build & Dev Tools
| Tool | Use |
|---|---|
| Bazel | Primary build system |
| Git | Version control |
| Docker | Containerization |
| Terraform | Infrastructure as code |
| Kubernetes / EKS | Orchestration |
Data Engineering
| Tool | Use |
|---|---|
| Apache Spark / Databricks | Large-scale data processing |
| Apache Airflow | Workflow orchestration |
| Kafka / Kinesis | Streaming data |
| Ray | Distributed compute |
Frontend / Visualization
| Tool | Use |
|---|---|
| React (primary), Vue.js, Angular | Web frontends for internal tools |
| three.js / Babylon.js | 3D rendering |
| Vulkan / OpenGL | 3D graphics APIs |
Monitoring & Observability
| Tool | Use |
|---|---|
| Prometheus + Grafana | Metrics and dashboards |
| NVIDIA DCGM Exporter | GPU-level monitoring |
| CloudWatch | AWS 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
| System | Redundancy Approach |
|---|---|
| Steering | Dual steering platforms (primary + backup) for bidirectional control |
| Braking | Multiple backup functions, multi-technology approach, third independent emergency brake |
| Power | Two independent battery packs + redundant power converters + two additional 12V backup batteries |
| Compute | Dual mirrored computer systems in vehicle floor with cross-verified logic domains |
| Sensors | 360° overlapping FOV; operational if individual sensors fail |
| Connectivity | Three cellular modems operating simultaneously with load balancing |
| Autonomy | Main driving stack + independent Collision Avoidance System |
| Drivetrain | Two 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)
| Attribute | Detail |
|---|---|
| Size | 220,000 sq ft |
| Capacity | 10,000+ 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 |
| Approach | Modular — major components pre-assembled by suppliers; Zoox does final integration |
| Automation | Robots handle adhesive dispensing, AGVs transport vehicles; humans do most assembly |
| Notable | No welding, cutting, or painting on-site — low power draw vs. traditional auto plants |
End-of-Line Testing Sequence
- Sensor calibration bay — automated turntable, halogen lights, radar targets; calibrates LiDAR, cameras, IR, radar
- Wheel & headlight alignment — both ends (bidirectional); active suspension calibration; electronic steering zeroed
- Dynamometer testing — autonomous stress tests up to 75 mph; laser-based lateral drift correction
- Water leak / rain simulation — validates seals and sensor cleaning system (water spray + air blasts)
- Light tunnel & appearance — exterior lights, doors, speakers, touchscreens
- Factory Static Test — VIN verification, safety faults, seatbelts, emergency release, two-way audio, NHTSA label
- Buzz/Squeak/Rattle on outdoor test track
- Factory Dynamic Test — extended autonomous closed-loop driving (hours), both clockwise and counterclockwise
Fleet Operations & Teleoperation
Three Operations Teams (Foster City HQ)
| Team | Role |
|---|---|
| Mission Operations | "Air traffic controllers" — fleet health (tire pressure, battery, cabin/coolant temps), rerouting around street closures, demand-based fleet reallocation, weather advisories |
| TeleGuidance | Not 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 Support | Full 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
| Status | Cities |
|---|---|
| Public service | Las Vegas (Sep 2025), San Francisco (Nov 2025) |
| Testing | Austin, Miami, Seattle, Atlanta, Los Angeles, Washington D.C., Columbus OH |
| Next launches | Austin, Miami |
Fleet size: ~50 purpose-built robotaxis + hundreds of Toyota Highlander test vehicles.
Regulatory & Safety Record
NHTSA Exemptions
| Exemption | Detail |
|---|---|
| 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
| Recall | Date | Vehicles | Issue | Remedy |
|---|---|---|---|---|
| 25E-019 | Mar 2025 | 258 | Over-cautious hard braking for bicyclists near crosswalks; incorrect collision anticipation from rear | OTA update to v24.32 |
| 25E-037 | May 2025 | 270 | At <0.5 m/s, failure to detect prone VRU immediately adjacent | OTA update (deployed May 21, 2025) |
| 25E-090 | Dec 2025 | 332 | Unnecessary 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
| Period | Registered Vehicles | Miles (safety driver) | Driverless Miles | Disengagements | Miles/Disengagement |
|---|---|---|---|---|---|
| 2023 (Dec 2022 – Nov 2023) | 281 | 710,409 | 11,263 | 4 | ~177,602 |
| 2024 (Dec 2023 – Nov 2024) | 380 | 951,871 | 37,804 | — | — |
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
| Partner | Role |
|---|---|
| Amazon | Parent company (acquired 2020, ~$1.3B); AWS infrastructure |
| NVIDIA | DRIVE platform GPUs (on-vehicle + data center); partnership since 2017 |
| Hesai Technology | LiDAR sensors (multi-year partnership) |
| Teledyne FLIR | Longwave infrared (thermal) cameras |
| Panasonic Energy | 2170-format Li-ion battery cells (multi-year deal, starting early 2026) |
| ZF | Four-wheel steering, chassis modules, occupant safety/protection systems |
| Intel | Xeon CPUs (4 per vehicle); Movidius VPU in FLIR modules |
| Microchip | Automotive chips |
| Formula One Williams Racing | Multi-season partnership (simulation expertise sharing) |
| Las Vegas Golden Knights | Multi-year collaboration |
| Strio.AI | Acquired 2022 for robotics/AI automation (Boston R&D) |
Competitive Comparison
| Dimension | Zoox | Waymo | Tesla | Cruise |
|---|---|---|---|---|
| Vehicle | Purpose-built, no steering wheel | Retrofitted Jaguar I-Pace | Modified Model Y (CyberCab planned) | Retrofitted Chevy Bolt EV (paused) |
| Sensors | LiDAR + radar + cameras + LWIR + mics (~64) | 13 cameras, 4 LiDARs, 6 radars, audio | Camera-only (Tesla Vision) | LiDAR + radar + cameras |
| Unique sensor | LWIR thermal cameras (first in industry) | — | — | — |
| Autonomy level | L4 geo-fenced | L4 geo-fenced | L4 geo-fenced (Austin pilot) | L4 geo-fenced (paused) |
| Operational cities | 2 (LV, SF) | 4+ (Phoenix, LA, SF, Austin) | 1 (Austin pilot) | Paused after 2023 incident |
| Fleet size | ~50 robotaxis + hundreds of test vehicles | 700+ | Consumer vehicles | GM-owned fleet (paused) |
| Pricing | Free (paid 2026) | ~$20/trip | ~$4/ride (Austin) | N/A |
| Backing | Amazon | Alphabet | Tesla | GM (scaled back) |
| Perception | Triple-redundant (AI + geometric + safety net) | Multi-modal fusion | Neural net vision-only | Multi-modal fusion |
| Vehicle design | Full vertical integration | Retrofit + custom sensor pod | Mass-market vehicle + software | Retrofit |
Key Zoox Differentiators
- Only company with a purpose-built vehicle (no retrofit compromises)
- Only company using LWIR thermal cameras in the sensor stack
- Triple-redundant perception (AI + geometric + safety net)
- Bidirectional driving eliminates three-point turns
- Full vertical integration from vehicle design through fleet operations
- Amazon/AWS infrastructure advantage for simulation and data processing
- Fail-operational (not just fail-safe) — aviation-inspired redundancy
Research & Publications
| Publication | Venue | Topic |
|---|---|---|
| Scenario Diffusion | NeurIPS 2023 | Controllable driving scenario generation with latent diffusion models; autoencoder + diffusion architecture; ~1 sec/scenario on single GPU |
| Sensor Staleness Framework | Amazon Science 2025 | Model-agnostic temporal feature engineering; pedestrian precision 2×, recall 6× improvement |
| QTP (Query-centric Trajectory Prediction) | Amazon Science | Data-driven behavior prediction replacing UAP |
| CLAMS Localization | Amazon Science | Infrastructure-free calibration using natural features |
| Foundation Model for AV | AWS 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
| Division | Key Leaders |
|---|---|
| Autonomy Software | VP: Marc Wimmershoff |
| Perception | Director: Bat-El Shlomo; also Ruijie He (Strio.AI) |
| Prediction | — |
| Planning & Control | — |
| Simulation | Scenario frameworks, 3D sensor sim, sim data platform, sim infra |
| ML Infrastructure | CPU/GPU resources, HPC |
| Driving Tools | Real-time operational tools, 3D visualization |
| Developer Platforms | — |
| SDMA (Systems Design & Mission Assurance) | V&V, safety metrics, safety clearance |
| Firmware | Embedded systems |
| Manufacturing Operations | Hayward and Fremont facilities |
| Safety Strategy & Operations | Standards compliance, safety cases |
| Homologation | Regulatory filings |
| User Experience | App, in-vehicle experience |
Locations: Foster City (HQ), San Francisco, Las Vegas, Boston (R&D), San Diego.
Sources
Official Zoox
- zoox.com — Homepage, /about, /vehicle, /autonomy, /safety, /careers
- Zoox Journal — Technical articles on redundancy, planner, perception, end-of-line testing, operational safety, mapping, prediction
- Zoox Vehicle Brochure (PDF)
Amazon Science
AWS & NVIDIA
- AWS re:Invent 2025 AMZ304
- Zoox AWS Case Study
- NVIDIA Blog: Zoox Autonomous Robotaxi
- NVIDIA Blog: Zoox Ride-Hailing
Regulatory
- NHTSA Recall 25E-019
- NHTSA Recall 25E-037
- NHTSA Recall 25E-090
- NHTSA Demonstration Exemption Press Release
- Federal Register: FMVSS Exemption (Docket NHTSA-2025-0523)
- California DMV Disengagement Reports
Infrastructure
Trade Press
- TechCrunch: Zoox production facility
- TechCrunch: Zoox crash safety
- Electrek: Zoox battery
- InsideEVs: Robotaxi first ride
- Assembly Magazine: Factory
- CNBC: Zoox robotaxi production
- The Robot Report: Simulation
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.