Waymo: Comprehensive Autonomous Vehicle Technology Stack

Last updated: March 2026

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
Fleet Operations & Ride-Hailing
Regulatory & Safety Record
Key Partnerships & Suppliers
Research & Publications
Engineering Organization
Competitive Differentiators
Open Source & Datasets

1. Company Overview

Corporate Identity

Attribute	Detail
Legal Name	Waymo LLC
Parent Company	Alphabet Inc. (majority owner)
Headquarters	Mountain View, California, USA
Founded	January 17, 2009 (as Google Self-Driving Car Project); spun out as Waymo in December 2016
Employees	~3,776 (as of early 2026)
Valuation	$126 billion (post-money, February 2026)
Annual Revenue Run Rate	~$350 million (late 2025)
Co-CEOs	Dmitri Dolgov (technology), Tekedra Mawakana (business)

Funding History

Round	Date	Amount Raised	Post-Money Valuation	Lead Investors
Series A	March 2020	$2.25B (grew to $3.2B)	~$30B	Silver Lake, Mubadala, Magna, Andreessen Horowitz
Series B	June 2021	$2.5B	~$30B	Alphabet and external investors
Series C	October 2024	$5.6B	$45B	Alphabet, Andreessen Horowitz, Silver Lake, Fidelity, Tiger Global, T. Rowe Price
Series D	February 2026	$16B	$126B	Dragoneer, DST Global, Sequoia Capital, Alphabet
Total Raised		~$27.1B

Key Milestones

Year	Milestone
2009	Google Self-Driving Car Project launched at Google X by Sebastian Thrun
2010	Project publicly revealed by the New York Times (Oct 9)
2012	First autonomous vehicle licensed in Nevada (modified Prius)
2015	First fully driverless ride on public roads (Austin, TX; no steering wheel)
2016	Spun out from Google X as "Waymo" under Alphabet
2017	Launched Early Rider Program in Phoenix metro with Chrysler Pacificas
2018	Launched Waymo One commercial ride-hailing in Phoenix (with safety drivers)
2019	Surpassed 10 billion simulated miles
2020	Fully driverless (rider-only) rides began in Phoenix; 5th-gen hardware unveiled
2021	Rider-only rides in San Francisco begin; co-CEO structure established
2022	Expanded rider-only operations in SF; began Waymo One rides in SF
2023	Waymo Via trucking unit closed to focus on robotaxis; retired Chrysler Pacifica fleet
2024	Launched in Los Angeles; 6th-gen Waymo Driver announced; surpassed 100K weekly paid rides
2025	Launched in Austin and Atlanta; 450K+ weekly rides; 2,500 robotaxis in service; Magna factory opened in Mesa, AZ
2026	6th-gen Waymo Driver begins autonomous operations; $16B raised; Waymo World Model unveiled; 200M autonomous miles logged; expansion to 10+ cities

2. Vehicle Platform

Vehicle Generations

Vehicle	Partner OEM	Generation	Status	Battery	Architecture	Approx. Fleet Count
Toyota Prius	Toyota	Gen 1-2	Retired (2009-2012)	Hybrid	N/A	~12
Lexus RX450h	Toyota/Lexus	Gen 3	Retired (2012-2017)	Hybrid	N/A	~30
Chrysler Pacifica Hybrid	FCA/Stellantis	Gen 4	Retired (2017-2023)	16 kWh PHEV (33 mi EV range)	N/A	~1,000
Jaguar I-PACE	Jaguar Land Rover	Gen 5	Active	90 kWh Li-ion, 234 mi EPA range	400V	~1,500+ (scaling to 3,500 by end 2026)
Waymo Ojai (Zeekr RT)	Geely/Zeekr	Gen 6	Active (Feb 2026)	93 kWh LFP, 800V	800V	Initial deployments; ramping through 2026

Jaguar I-PACE (5th Generation Waymo Driver)

The all-electric Jaguar I-PACE has served as Waymo's primary commercial vehicle since 2020. Key specifications:

Drivetrain: Dual-motor AWD, 394 hp combined
Battery: 90 kWh liquid-cooled LG Chem lithium-ion
EPA Range: 234 miles (377 km)
Charging: 100 kW DC fast charge (0-80% in ~40 min)
Body: Aluminum-intensive monocoque
Sensor Hardware: 5th-gen Waymo Driver with 29 cameras, 5 LiDAR, 6 radar, external audio receivers
Sensor Pod: Roof-mounted pod with spinning LiDAR (Laser Bear Honeycomb evolution), plus front/rear bumper pods containing cameras and radar. The pod design features overlapping fields of view for full 360-degree coverage.
Waymo received its final Jaguar I-PACE delivery and plans to retrofit 2,000+ additional I-PACE vehicles through 2026 at the Magna Mesa factory.

Waymo Ojai / Zeekr RT (6th Generation Waymo Driver)

The purpose-built robotaxi developed with Geely's Zeekr brand represents the first mass-produced vehicle designed from the ground up for autonomous ride-hailing.

Official Name: Waymo Ojai (named after the California city; internally "Zeekr RT")
OEM Partner: Zeekr (majority owned by Geely, parent of Volvo Cars)
Design Origin: Designed in Sweden, manufactured in China (Zeekr), integrated in Mesa, AZ (Magna/Waymo)
Drivetrain: Rear-mounted single motor, 268 hp
Battery: 93 kWh lithium iron phosphate (LFP)
Architecture: 800V electrical architecture (enabling rapid DC fast charging for 24/7 fleet operation)
Doors: Three sliding doors (one driver-side conventional, two sliding passenger doors)
Interior: Purpose-built for passengers -- five seats, center tunnel table, no traditional driver controls in rider cabin
Sensor Hardware: 6th-gen Waymo Driver with 13 cameras, 4 LiDAR, 6 radar, external audio receivers (EARs)
Sensor Pod: Smaller, more integrated sensor package with miniature wipers for camera and LiDAR lens cleaning. Front and rear sensor pods contain cameras, radar, and LiDAR units distributed to maintain 360-degree overlap.
Distinguishing Feature: First mass-produced vehicle purpose-built for autonomous driving (per Zeekr CEO Andy An)

Chrysler Pacifica Hybrid (Retired)

Partnership: Began May 2016 with FCA (now Stellantis)
Fleet Size: Initially 100 units (Dec 2016), expanded to 600 by 2018, orders placed for up to 62,000
Actual Deployments: ~1,000 converted vehicles served riders before retirement in 2023
Modifications: Custom electrical, powertrain, chassis, and structural engineering for full autonomy
Reason for Retirement: Transition to all-electric fleet (Jaguar I-PACE)

3. Sensor Suite

Generation Comparison

Sensor Type	5th Gen (I-PACE)	6th Gen (Ojai)	Change
Cameras	29	13	-55%
LiDAR	5	4	-20%
Radar	6	6	0%
External Audio Receivers	Yes	Yes	--
Max Detection Range	>500 m	>500 m	--
Field of View	360 deg overlapping	360 deg overlapping	--

LiDAR (In-House Design)

Waymo designs and manufactures its own LiDAR sensors, making it one of the only AV companies with fully vertically integrated LiDAR.

Types of LiDAR (3 distinct categories):

Long-Range LiDAR (Roof-Mounted)
- Custom "Laser Bear Honeycomb" design evolution
- 360-degree rotation
- Range: >300 meters
- The 5th-gen version featured a 95-degree vertical field of view (VFOV)
- 6th-gen version: reengineered illumination and internal data processing to better penetrate weather and reduce point cloud distortion near highly reflective signs
Short-Range / Perimeter LiDAR
- Positioned at four points around the vehicle sides (5th gen) or optimized placements (6th gen)
- Provides uninterrupted surround view close to the vehicle body
- Critical for detecting vulnerable road users (pedestrians, cyclists) at close range
- Centimeter-scale range accuracy
High-Resolution Forward LiDAR
- First-of-its-kind zooming capability to dynamically focus on distant objects
- Provides detailed point clouds at extended ranges

Cost Reduction: Waymo reduced LiDAR unit cost by more than 90% through in-house design (from $75,000+ per unit to an estimated sub-$7,500).

Core Components: Designed and built in California. Custom-designed chips and optical elements.

Cameras

5th Generation:

29 cameras providing 360-degree coverage with overlapping fields of view
Capable of identifying stop signs at >500 meters

6th Generation:

13 cameras (reduced count, higher capability per unit)
Next-generation 17-megapixel imager
Exceptional thermal stability across automotive temperature ranges
Day/night operation with overlapping fields of view up to 500 meters
Miniature wipers on front sensor pod to clear debris

Radar (In-House Imaging Radar)

Waymo developed one of the world's first automotive imaging radar systems.

Count: 6 radar units (both 5th and 6th gen)
Type: Custom in-house imaging radar
Capabilities:
- Dense temporal maps tracking distance, velocity, and size of objects
- Continuous 360-degree view
- Tracks targets at >500 meters
- Operates in all lighting and weather conditions (rain, fog, snow, direct sunlight)
- Detects static, barely moving, and fully moving objects
- Higher resolution and enhanced signal processing vs. conventional automotive radar
6th-gen Improvements: New in-house algorithms for improved rain/snow performance; added dense temporal mapping for real-time object speed, size, and trajectory tracking

External Audio Receivers (EARs)

Array of microphones positioned around the vehicle exterior
Detects and localizes emergency vehicle sirens, honking, construction noise
Enables the Waymo Driver to respond to auditory cues that cameras and LiDAR cannot capture

4. Onboard Compute

Compute Architecture

Waymo designs its own onboard compute platform, treating it as a tightly integrated system with its sensor suite.

Component	Details
Primary Compute	Custom board combining server-grade CPUs and GPUs
Secondary Compute	Redundant backup computer running continuously in parallel
CPUs	Intel Xeon processors (historically confirmed)
GPUs	NVIDIA GPUs (likely A-series or successor for inference)
FPGAs	Intel Arria FPGAs for low-latency sensor preprocessing
Connectivity	Intel XMM LTE/5G modems; Intel Gigabit Ethernet for internal bus
Sensor Data Throughput	~20 GB/s aggregate from all sensors
End-to-End Latency	~3 ms (sensor input to actuation command)
Redundancy	Full dual-compute architecture; secondary system can bring vehicle to safe stop
Design	Entirely in-house by Waymo hardware engineers

Redundancy Philosophy

Dual compute boards: Primary and secondary systems run simultaneously; failover is seamless
Independent power supplies: Separate power paths for primary and backup systems
Watchdog monitoring: Hardware watchdogs detect compute failures within milliseconds
Graceful degradation: If primary compute fails, backup system executes a minimal risk condition (MRC) -- typically pulling over and stopping safely

5. Autonomy Software Stack

The Waymo Driver software is organized as a multi-stage pipeline, though recent developments are pushing toward end-to-end learned approaches.

Pipeline Architecture

Sensors --> Perception --> Prediction --> Planning --> Control --> Actuators
               |               |              |            |
               v               v              v            v
         3D Detection    Behavior        Trajectory    Steering/
         Tracking        Forecasting     Generation    Braking/
         Segmentation    Intent          Optimization  Throttle
         Classification  Recognition     Risk Eval

5.1 Perception

The perception module fuses data from all sensor modalities (LiDAR, camera, radar, audio) to build a comprehensive understanding of the environment.

Key capabilities:

3D Object Detection: Detects 100+ object classes including vehicles, pedestrians, cyclists, scooters, animals, traffic cones, construction equipment
Semantic Segmentation: Per-point LiDAR segmentation and per-pixel camera segmentation
Object Tracking: Temporal association of detected objects across frames with globally unique tracking IDs
Lane Detection & Road Graph: Real-time identification of lane boundaries, crosswalks, curbs
Traffic Signal Recognition: Detection and classification of traffic lights (including arrow signals), stop signs, speed limit signs
Free Space Estimation: Drivable area computation

Model architectures:

Transformer-based models (ViTs) for camera-based perception
BEV (Bird's Eye View) representations via BEVFormer-style architectures
PointPillars-derived architectures for LiDAR point cloud processing
Multi-modal fusion networks combining LiDAR + camera features
Neural Architecture Search (NAS) for optimized model design: 10,000+ architectures explored, yielding 10% lower error rate at same latency or 20-30% faster inference at same quality

5.2 Prediction

The prediction module forecasts the future trajectories and behaviors of all detected agents.

Key capabilities:

Multi-agent trajectory prediction over 8-11 second horizons
Joint prediction of 100+ agents simultaneously
Intent recognition (lane change, turn, stop, yield)
Interaction modeling between agents (e.g., merging, negotiating intersections)

Model architectures:

Graph Neural Networks (GNNs) for modeling agent-to-agent and agent-to-road interactions (VectorNet)
MultiPath/MultiPath++: Fixed anchor trajectory hypotheses with Gaussian mixture outputs
Temporal graph networks and diffusion models for motion forecasting
Scaling laws applied: 10x data/compute yields 20-30% error reduction in forecasting

5.3 Planning

The planning module generates safe, comfortable, and efficient trajectories.

Key capabilities:

Trajectory generation with spatiotemporal optimization
Risk-aware decision making under uncertainty
Goal fulfillment (follow route, reach destination)
Negotiation behaviors (unprotected left turns, merges, yielding)
Comfort optimization (smooth acceleration, deceleration, steering)

Approach evolution:

Historical: Rule-based + optimization-based planning
Current: ML-driven motion planning using foundation models, shifting from hand-crafted rules to learned policies
Scaling laws demonstrate power-law gains in trajectory quality with more data and compute

5.4 Control

The control module translates planned trajectories into precise actuator commands.

Key capabilities:

Model Predictive Control (MPC) hybridized with neural policies
Reinforcement learning-derived control policies
Real-time adaptation to road surface conditions, wind, vehicle load
Smooth, human-like driving behavior

5.5 End-to-End Trends

Waymo is actively developing end-to-end approaches that collapse the traditional pipeline:

EMMA (End-to-End Multimodal Model): Maps raw camera data directly to planner trajectories, perception outputs, and road graph elements using a single model built on Gemini
The production Waymo Driver now uses a foundation model trained end-to-end, similar in philosophy to approaches by Tesla and Wayve
Co-training across perception, detection, and planning tasks improves all three domains simultaneously

6. Machine Learning & AI

Foundation Models

EMMA (End-to-End Multimodal Model for Autonomous Driving)

Attribute	Detail
Published	October 2024 (arXiv: 2410.23262); updated September 2025
Base Model	Google Gemini (multimodal large language model)
Input	Raw camera sensor data, navigation instructions, ego vehicle status
Output	Planner trajectories, 3D object detections, road graph elements
Representation	All non-sensor inputs/outputs encoded as natural language text
Multi-task	Co-trained on planning, object detection, and road graph tasks
Performance	State-of-the-art motion planning on nuScenes; competitive on WOMD and WOD
Limitations	Processes limited image frames; no LiDAR/radar input; computationally expensive
Venue	ICLR 2025

Waymo Foundation Model (Production)

Single large foundation model powering the production Waymo Driver
Trained on Google TPU pods (same infrastructure used for Gemini)
"Generalizable" -- designed to transfer across vehicle platforms and geographies
Built on JAX framework (same as Google DeepMind's Gemini)

Training Infrastructure

Component	Detail
Framework	JAX (primary), TensorFlow (legacy/supplementary)
Hardware	Google TPU pods (v4, v5e, likely v6/Trillium); NVIDIA GPUs for some workloads
TPU Efficiency	Up to 15x more efficient training vs. GPU baselines
Distributed Training	Custom libraries for job scheduling, resource management, data distribution, model synchronization
Training Data	Hundreds of millions of real-world driving miles + billions of simulated miles
Data Processing	Petabytes of complex sensor data processed through scalable pipelines
AutoML/NAS	Reinforcement learning and random search over 10,000+ architectures; final selection based on accuracy and inference cost
Shared Infrastructure	Optimizations from Gemini training directly portable to Waymo Driver training

Model Architecture Components

Architecture	Application	Details
Vision Transformers (ViTs)	Camera perception	Self-attention over image patches for object detection and segmentation
BEVFormer	Multi-camera 3D perception	Generates bird's-eye-view feature maps from surround cameras
PointPillars	LiDAR 3D detection	Point cloud encoded into vertical pillars, processed by 2D CNN; runs at 62 Hz
VectorNet	HD map & agent encoding	Hierarchical GNN operating on vectorized map and trajectory representations
MultiPath / MultiPath++	Behavior prediction	Fixed anchor trajectories with Gaussian mixture outputs; state-of-the-art on WOMD
Diffusion Models	Motion forecasting & simulation	Used in SceneDiffuser for traffic simulation initialization and rollout
Gemini-based LLM	End-to-end driving (EMMA)	Multimodal LLM mapping sensor data to driving outputs via natural language

AutoML and Neural Architecture Search

Waymo collaborates with Google Brain/DeepMind to apply AutoML to autonomous driving:

Process: Explore 10,000+ architectures using RL and random search; pre-select 100 candidates; choose 1 winner based on accuracy and inference cost
Applications: LiDAR semantic segmentation, traffic lane detection/localization
Results: Architectures with 10% lower error rate at same latency, or 20-30% faster at same quality
NAS Cells: Family of neural network building blocks (NAS cells) composed to build task-specific architectures

7. Mapping & Localization

HD Map Creation

Waymo relies on pre-built High-Definition (HD) maps as a core component of its autonomy stack.

Map Creation Process:

Manual Survey Drives: Sensor-equipped vehicles manually driven through every street in a new operational domain
3D LiDAR Reconstruction: Custom LiDAR paints a detailed 3D picture of the environment
Annotation: Maps enriched with semantic information -- lane boundaries, traffic signals, stop signs, crosswalks, speed limits, curb heights, driveway locations
Validation: Automated and manual quality checks ensure centimeter-level accuracy
Updates: Maps continuously updated as the fleet detects changes in the environment

Map Specifications

Attribute	Detail
Accuracy	Centimeter-level (cm-level)
Content	3D geometry, lane-level topology, traffic control devices, crosswalks, speed zones, construction zones
Format	Proprietary vectorized representation
Coverage	All operational cities fully mapped before service launch
Update Frequency	Continuous, driven by fleet observations and dedicated mapping runs

Localization

Primary Method: LiDAR-based localization against pre-built HD point cloud maps
Supplementary: GPS/GNSS, IMU, wheel odometry
Accuracy: Centimeter-level position accuracy in 6-DOF (position + orientation)
Robustness: Multi-sensor fusion ensures localization even when individual sensors are degraded (e.g., GPS in urban canyons)
SLAM Integration: LiDAR SLAM used during map creation; real-time localization uses scan-matching against pre-built maps with misalignment correction

8. Simulation Platform

Overview

Waymo's simulation infrastructure is among the most extensive in the AV industry, with over 20 billion simulated miles driven (vs. 200+ million real-world autonomous miles).

SimulationCity

Waymo's primary simulation platform, unveiled in July 2021.

Feature	Detail
Scale	Up to 25,000 virtual Waymo vehicles driving up to 10 million miles per day
Fidelity	Recreates raindrops on sensors, dimming light, solar glare, road surface variations
Data Sources	20M+ real autonomous miles, NHTSA Crash Data Systems, Naturalistic Driving Study Data
Scenario Types	Replay of real-world encounters, procedural generation, adversarial testing, regression testing
Capabilities	New vehicle platform bring-up, geographic validation, operational refinement
Statistical Modeling	Draws random outcomes from statistical distributions of real-world behaviors

SurfelGAN

Published: CVPR 2020
Purpose: Generate realistic sensor data for novel viewpoints from limited LiDAR + camera data
Method: Texture-mapped surfels for efficient scene reconstruction preserving 3D geometry, appearance, and scene conditions
Output: Photorealistic camera images for novel vehicle positions and orientations
Application: Data augmentation, counterfactual scenario generation

Waymo World Model (February 2026)

Waymo's most advanced simulation system, built on Google DeepMind's Genie 3.

Feature	Detail
Foundation	Google DeepMind's Genie 3 (general-purpose world model for photorealistic 3D environments)
Output	Multi-sensor: generates both camera AND LiDAR data simultaneously
Control Modes	(1) Driving action control, (2) Scene layout control, (3) Text prompt control
Rare Event Generation	Can simulate tornadoes, flooded streets, elephants on the road, and other near-impossible scenarios
Data Conversion	Converts ordinary dashcam or cell phone videos into full multimodal (camera + LiDAR) simulations
3D Transfer	Pre-trained on vast 2D video corpus, then post-trained to produce 3D LiDAR outputs matching Waymo's hardware
Use Case	Every dashcam incident on the internet becomes a potential training scenario

SceneDiffuser / SceneDiffuser++

SceneDiffuser: Scene-level diffusion model for traffic simulation initialization and rollout (NeurIPS 2024). Achieves 16x fewer inference steps via amortized denoising. Top open-loop and best closed-loop diffusion performance on Waymo Open Sim Agents Challenge.
SceneDiffuser++: City-scale traffic simulation via generative world model. First end-to-end generative world model capable of point A-to-B simulation at city scale.

Simulation Scale

Metric	Value
Total Simulated Miles	>20 billion
Daily Simulated Miles	Up to 10 million
Virtual Fleet Size	Up to 25,000 simultaneous virtual vehicles
Closed-Course Scenarios	>40,000 unique scenarios
Simulated Driving Per Year	Billions of miles

9. Cloud & Data Infrastructure

Google Cloud Integration

As an Alphabet subsidiary, Waymo has deep integration with Google Cloud infrastructure.

Component	Detail
Compute	Google TPU pods (v4, v5e, Trillium/v6) for training; GPU clusters for inference development
Storage	Google Cloud Storage (GCS) for petabyte-scale sensor data
Data Warehouse	BigQuery for structured analytics and metrics
ML Platform	Vertex AI and custom training orchestration
Networking	High-bandwidth interconnect between TPU pods (9.6 Tb/s inter-chip, per Ironwood specs)
HBM Per Superpod	Up to 1.77 PB shared High Bandwidth Memory (9,216-chip Ironwood superpod)

Data Pipeline

Data Volume: Many petabytes of complex sensor data (LiDAR point clouds, camera images, radar returns, audio)
Pipeline Function: Raw sensor data ingestion, preprocessing, annotation, training data generation, model evaluation
Auto-Labeling: High-accuracy 3D auto-labeling system generates 3D bounding boxes for road agents
Scale: Data volume "exploding" as Waymo scales to new cities and vehicle platforms
Annotation Tools: Combination of automated labeling (ML-based) and human annotation for ground truth

Data Flow Architecture

Fleet Vehicles (20 GB/s per vehicle)
    |
    v
Cellular Upload (prioritized data selection)
    |
    v
Google Cloud Storage (petabyte-scale raw data)
    |
    v
Data Processing Pipeline (preprocessing, auto-labeling, quality checks)
    |
    v
Training Data Store (curated datasets for ML training)
    |
    v
TPU Pods (distributed training)
    |
    v
Model Registry (versioned models, A/B testing)
    |
    v
OTA Deployment to Fleet

10. Programming Languages & Tools

Programming Languages

Language	Primary Use
C++	Core autonomy stack (perception, planning, control), real-time onboard software, sensor drivers, low-latency inference
Python	ML model development, training scripts, data analysis, prototyping, tooling
Go	Backend services, fleet management, cloud infrastructure, APIs
Java	Some backend services (Alphabet ecosystem compatibility)
Starlark	Build configuration language (Python dialect for Bazel BUILD files)

Build System

Tool	Detail
Bazel	Primary build system (open-source version of Google's internal Blaze). Handles multi-language monorepo builds across C++, Python, Go, Java. Hermetic builds ensure reproducibility.
Blaze	Google-internal predecessor to Bazel; Waymo likely uses Blaze internally given Alphabet's infrastructure
Key Benefits	Incremental builds, remote caching, remote execution, cross-platform compilation, deterministic outputs

Internal Tools and Frameworks

Tool	Purpose
Custom ML Libraries	Enhance TensorFlow and JAX for Waymo-specific scalability, reliability, and performance
Training Orchestration	Custom distributed training infrastructure: job scheduling, resource management, data distribution, model synchronization
Simulation Framework	SimulationCity and Waymo World Model toolchain
Data Labeling	Auto-labeling systems + annotation management tools
Fleet Management	Proprietary vehicle dispatch, routing, monitoring, and remote assistance systems
Map Toolchain	HD map creation, validation, and update pipeline
CI/CD	Continuous integration and deployment for both onboard software and cloud services

ML Frameworks

Framework	Usage
JAX	Primary ML framework (rebuilt entire training stack on JAX); same framework as Google DeepMind's Gemini
TensorFlow	Legacy training pipelines and some production inference
Flax/Haiku	Neural network libraries on top of JAX
XLA	Compiler for optimizing TPU/GPU execution

11. Safety & Redundancy Architecture

Safety Methodology

Waymo employs a multi-layered safety approach organized across three technology layers.

Three Layers of Safety:

Layer	Scope	Methods
Hardware (Architecture)	Sensors, compute, actuators	Redundant sensors, dual compute, backup braking/steering, DFMEA, FTA
ADS Behavior (Software)	Perception, prediction, planning, control	STPA, scenario-based verification, simulation testing, ML robustness
Operations	Fleet management, remote assistance, maintenance	Operational Design Domain (ODD) constraints, geo-fencing, remote support

Hazard Analysis Methods:

STPA (Systems-Theoretic Process Analysis): Identifies hazards from control structure interactions
FTA (Fault Tree Analysis): Top-down deductive analysis of failure paths
DFMEA (Design Failure Modes and Effects Analysis): Systematic evaluation of component failure modes
Custom Software Safety Analysis: Waymo-specific safety analysis for ML-based systems

Hardware Redundancy

System	Redundancy Approach
Compute	Dual onboard computers running simultaneously; backup can bring vehicle to safe stop
Braking	Redundant braking actuators (specified as necessary for driverless operation)
Steering	Redundant steering actuators
Power	Independent power paths for primary and backup systems
Sensors	Overlapping fields of view across LiDAR, cameras, and radar; any single sensor failure covered by remaining sensors
Communication	Multiple cellular connections for fleet management and remote assistance

Testing Portfolio

Test Type	Scope
Simulation	Billions of miles; regression testing; rare event coverage; system-level
Closed-Course	>40,000 unique scenarios; controlled environment testing; hardware-in-the-loop
Public Road	200+ million autonomous miles; real-world validation
Unit Tests	Software component verification
Integration Tests	Subsystem interaction validation
Bench Tests	Hardware component testing
Hardware-in-the-Loop (HIL)	Sensor and compute hardware tested with simulated inputs

Safety Readiness Determination

Waymo uses a formal Safety Readiness Determination (SRD) process before deploying in new domains:

Hazard analysis and risk assessment
Design and development with built-in robustness
Scenario-based verification (simulation + closed-course + limited public road)
Operational readiness review
Phased deployment (safety drivers first, then driverless)

Reference: Waymo's Safety Methodologies and Safety Readiness Determinations (arXiv:2011.00054, published November 2020)

12. Fleet Operations & Ride-Hailing

Waymo One Service

Metric	Value (as of March 2026)
Weekly Paid Rides	450,000+ (Dec 2025); targeting 1M by end of 2026
Fleet Size	~2,500 vehicles (Nov 2025); scaling to 3,500+ I-PACEs + Ojai additions
Total Autonomous Miles	200+ million (fully autonomous, public roads)
Rider-Only Miles	96+ million (no human driver present)
App	Waymo One (iOS and Android); also available via Uber app in select cities

Operational Cities (as of March 2026)

City	Launch Date	App	Fleet Vehicle	Notes
Phoenix, AZ	Oct 2020 (driverless)	Waymo One	Jaguar I-PACE	Largest and most mature market
San Francisco, CA	Mar 2022 (driverless)	Waymo One	Jaguar I-PACE	Major urban deployment
Los Angeles, CA	2024	Waymo One	Jaguar I-PACE	Rapidly expanding service area
Austin, TX	Early 2025	Uber app	Jaguar I-PACE	Partnership with Uber for dispatch and fleet ops
Atlanta, GA	Jun 2025	Uber app	Jaguar I-PACE	65 sq mi initial area; Downtown to Buckhead
Miami, FL	Early 2026 (early access)	Waymo One	Jaguar I-PACE	Moove handles fleet operations
Houston, TX	2026 (announced Feb)	TBD	TBD	Initial select riders
Dallas, TX	2026 (announced Feb)	TBD	TBD	Rolling access
San Antonio, TX	2026 (announced Feb)	TBD	TBD	Rolling access
Orlando, FL	2026 (announced Feb)	TBD	TBD	Rolling access

Announced future cities: Washington D.C. (2026), Nashville, Detroit, Las Vegas, San Diego, Denver, and international targets including London and Tokyo.

Pricing Model

San Francisco:

Base fare: $9.52
Per mile: $1.66
Per minute: $0.30
Average ride cost: ~$20.43

Los Angeles:

Base fare: $5.37
Per mile: $2.50
Per minute: $0.32

Price Comparison (April 2025 data):

Service	Average Ride Cost	Premium vs. Lyft
Waymo	$20.43	+41%
Uber	$15.58	+8%
Lyft	$14.44	Baseline

Trend: The pricing gap is narrowing. By late 2025, Waymo's premium dropped to ~12.7% over Uber and ~27.3% over Lyft. Waymo's average cost dropped 3.62% while Uber and Lyft prices increased.

Subscription: $29.99/month subscription available, pays for itself in under five rides vs. standard fares.

Fleet Management Partners

Partner	Role	Markets
Waymo (direct)	Full fleet operations	Phoenix, SF, LA
Uber	Dispatch via Uber app, fleet management	Austin, Atlanta
Moove	Fleet operations, facilities, charging	Miami (planned), Phoenix
Avomo (formerly Moove Cars)	Fleet management for Uber-dispatched rides	Austin, Atlanta

13. Regulatory & Safety Record

Autonomous Miles and Crash Data

Metric	Value	Source
Total Autonomous Miles (public road)	200+ million	Waymo (Feb 2026)
Rider-Only Miles	96+ million	Waymo
NHTSA-Reported Incidents	1,429 (Jul 2021 - Nov 2025)	NHTSA SGO data
Reported Injuries	117	NHTSA
Fatalities	2 (Waymo vehicle involvement)	NHTSA

Peer-Reviewed Safety Comparisons

Study 1: 7.1 Million Rider-Only Miles (Kusano et al., 2024; published in Traffic Injury Prevention):

Crash Type	Waymo ADS (IPMM)	Human Benchmark (IPMM)	Reduction
Any Injury Reported	0.41	2.80	-85% (human rate 6.7x higher)
Airbag Deployment	Significant reduction	--	Statistically significant

Study 2: 56.7 Million Rider-Only Miles (Kusano et al., 2025; published in Traffic Injury Prevention):

First statistically significant comparison at the Suspected Serious Injury+ (SSI+) level
Statistically significant reductions in: Any-Injury-Reported, Airbag Deployment, and SSI+ outcomes
Disaggregated into 11 crash types for granular analysis

California DMV Disengagement Reports

Period	Miles Driven (CA)	Fleet Size (CA)	Notes
Dec 2022 - Nov 2023	5,872K km (3,670K mi)	~430 vehicles	Peak testing mileage
Dec 2023 - Nov 2024	3,823K km (2,390K mi)	1,035 vehicles	35% fewer miles; shift from testing to commercial ops
Dec 2024 - Nov 2025	Part of 9M+ mi industry total	--	Commercial operations dominate

Note: As Waymo transitions from testing to commercial operations, traditional disengagement reporting becomes less meaningful. Waymo's commercial driverless miles (rider-only) are reported through NHTSA's SGO framework rather than California DMV disengagement reports.

Regulatory Framework

Federal: Voluntary compliance with NHTSA Standing General Order (SGO) for incident reporting; no federal AV legislation as of March 2026
California: CPUC permit for commercial robotaxi operations; CA DMV autonomous vehicle testing permit
Arizona: Permissive regulatory environment; no specific AV legislation required for operation
Other States: Regulatory engagement underway in Texas, Florida, Georgia, and Washington D.C.

14. Key Partnerships & Suppliers

Vehicle Manufacturing Partners

Partner	Relationship	Vehicle/Component
Geely / Zeekr	OEM partner; designed and manufactures the Zeekr RT (Waymo Ojai)	6th-gen robotaxi platform
Jaguar Land Rover	OEM partner (winding down)	Jaguar I-PACE base vehicle
Magna International	Manufacturing partner; co-operates Mesa, AZ integration facility	Sensor integration, assembly
Stellantis (FCA)	Former OEM partner	Chrysler Pacifica Hybrid (retired)

Ride-Hailing & Fleet Partners

Partner	Relationship	Details
Uber	Ride-hailing platform integration	Waymo rides available via Uber app in Austin and Atlanta
Moove	Fleet management	Operations, facilities, and charging in Miami; fleet ops in Phoenix
Avomo	Fleet management	Depot operations for Uber-dispatched markets

Autonomous Trucking Partners

Partner	Relationship	Status
Daimler Trucks	Strategic partnership for L4 autonomous trucks	Active; Waymo provides autonomy stack for Freightliner Cascadia
J.B. Hunt	Freight testing partner	Waymo Via closed (2023), but Daimler partnership continues
C.H. Robinson	3PL testing partner	Testing completed
UPS	Freight trial partner	Trial runs in Texas completed
Uber Freight	Freight platform integration	Exploratory partnership

Technology Partners

Partner	Relationship
Google DeepMind	Genie 3 world model foundation; shared research
Google Cloud	TPU infrastructure, storage, compute
Google Brain (now DeepMind)	AutoML/NAS collaboration for architecture search
Intel	Historical: Xeon CPUs, Arria FPGAs, XMM modems, Gigabit Ethernet

15. Research & Publications

Publication Statistics

Waymo maintains an active research program with publications at top-tier venues.

Year	Paper Count
2019	~8
2020	~15
2021	~12
2022	~14
2023	16
2024	13
2025	4+

Key Papers by Topic

Perception

Paper	Venue	Year	Key Contribution
PointPillars: Fast Encoders for Object Detection from Point Clouds (Lang et al.)	CVPR	2019	Pillar-based LiDAR encoding; 62 Hz real-time; 2-4x faster than prior methods
STINet: Spatio-Temporal-Interactive Network for Pedestrian Detection and Trajectory Prediction (Zhang et al.)	CVPR	2020	Joint pedestrian detection + trajectory prediction
Scalability in Perception for Autonomous Driving: Waymo Open Dataset (Sun et al.)	CVPR	2020	Introduced Waymo Open Dataset; benchmark for 3D detection
Depth Estimation Matters Most	--	2020	Per-object depth estimation for monocular 3D detection and tracking
DeepFusion: Lidar-Camera Deep Fusion for Multi-Modal 3D Object Detection	CVPR	2022	Cross-modal feature fusion for 3D detection

Behavior Prediction

Paper	Venue	Year	Key Contribution
MultiPath: Multiple Probabilistic Anchor Trajectory Hypotheses for Behavior Prediction (Chai et al.)	CoRL	2019	Fixed anchor trajectories with Gaussian mixture; one-pass inference
VectorNet: Encoding HD Maps and Agent Dynamics from Vectorized Representation (Gao et al.)	CVPR	2020	Hierarchical GNN on vectorized map/trajectory inputs
MultiPath++: Efficient Information Fusion and Trajectory Aggregation for Behavior Prediction (Varadarajan et al.)	ICRA	2022	State-of-the-art on WOMD and Argoverse
Large Scale Interactive Motion Forecasting for Autonomous Driving: The Waymo Open Motion Dataset (Ettinger et al.)	ICCV	2021	Introduced WOMD; 103K segments

End-to-End Driving

Paper	Venue	Year	Key Contribution
EMMA: End-to-End Multimodal Model for Autonomous Driving (Hwang et al.)	ICLR	2025	Gemini-based model mapping raw cameras to trajectories via natural language

Simulation

Paper	Venue	Year	Key Contribution
SurfelGAN: Synthesizing Realistic Sensor Data for Autonomous Driving (Yang et al.)	CVPR	2020	Surfel-based scene reconstruction + GAN for novel view synthesis
SceneDiffuser: Efficient and Controllable Driving Simulation	NeurIPS	2024	Diffusion-based traffic simulation; 16x fewer inference steps
SceneDiffuser++: City-Scale Traffic Simulation via a Generative World Model	--	2025	First end-to-end generative world model for city-scale simulation

Safety

Paper	Venue	Year	Key Contribution
Waymo's Safety Methodologies and Safety Readiness Determinations	arXiv: 2011.00054	2020	Formal safety framework: STPA, FTA, DFMEA, SRD process
Comparison of Waymo Rider-Only Crash Data to Human Benchmarks at 7.1M Miles (Kusano et al.)	Traffic Injury Prevention	2024	85% crash reduction vs. human benchmark
Comparison of Waymo Rider-Only Crash Rates at 56.7M Miles (Kusano et al.)	Traffic Injury Prevention	2025	First statistically significant SSI+ comparison

Conference Presence

Waymo regularly publishes at and sponsors:

CVPR (Computer Vision and Pattern Recognition)
NeurIPS (Neural Information Processing Systems)
ICRA (International Conference on Robotics and Automation)
ICLR (International Conference on Learning Representations)
CoRL (Conference on Robot Learning)
ICCV (International Conference on Computer Vision)
ECCV (European Conference on Computer Vision)
RSS (Robotics: Science and Systems)
IV (Intelligent Vehicles Symposium)

Patent Portfolio

Metric	Value
Total Patents Globally	3,476
Unique Patent Families	1,077
USPTO Applications	1,237
USPTO Grants	929
Grant Rate	97.07%
Active Patents	3,199
LiDAR Patents	626
Autopilot/Control Patents	186
Citation Impact	1.6x higher citation rate than Toyota; 2.3x higher than GM
Core R&D Themes	Autonomous vehicle, LiDAR sensor, light detection, trajectory planning

16. Engineering Organization

Leadership

Name	Title	Focus
Dmitri Dolgov	Co-CEO	Technology, engineering, and research. Former CTO and VP of Engineering. PhD in robotics; designed Stanley's perception stack (DARPA Grand Challenge). At the project since 2009.
Tekedra Mawakana	Co-CEO	Business operations, commercialization, public policy. Joined Waymo in 2017.
Dragomir (Drago) Anguelov	VP, Head of AI Foundations (Research)	Leads research team focused on ML for autonomous driving. Joined Waymo 2018. Distinguished Scientist.

Historical Key Figures

Name	Role	Contribution
Sebastian Thrun	Founding Lead	Stanford professor; led Project Chauffeur at Google X (2009-2013); built Stanley for DARPA Challenge
Chris Urmson	Engineering Lead	Led software development; CEO until departure in 2017; co-founded Aurora
Anthony Levandowski	Co-founder	Built first self-driving test vehicle; departed 2016; founded Otto (acquired by Uber); trade secret lawsuit
John Krafcik	CEO (2015-2021)	First dedicated CEO; led transition from project to company; launched Waymo One commercial service
Nathaniel Fairfield	VP, Engineering	Long-tenured engineering leader

Team Structure

Waymo's engineering organization is broadly organized into:

AI/ML Research (led by Drago Anguelov): Perception, prediction, planning, simulation research
Systems Engineering: Onboard software, hardware integration, real-time systems
Hardware: Sensor design (LiDAR, radar, cameras), compute platform, vehicle integration
Simulation & Testing: SimulationCity, Waymo World Model, closed-course testing
ML Infrastructure: Training infrastructure, data pipelines, model deployment
Fleet Operations Engineering: Remote assistance, dispatch optimization, fleet monitoring
Safety Engineering: System safety, hazard analysis, validation & verification
Mapping: HD map creation, localization, map maintenance

17. Competitive Differentiators

What Makes Waymo Unique

Differentiator	Detail
Longest Track Record	17+ years of continuous AV development (since 2009), more than any competitor
Most Autonomous Miles	200+ million real-world autonomous miles + 20+ billion simulated miles
Vertical Integration of Sensors	In-house LiDAR, radar, and camera design -- not reliant on third-party sensor companies
Google/Alphabet Ecosystem	Access to TPU pods, DeepMind research, Google Cloud, JAX/Gemini infrastructure
Commercial Operations at Scale	Only AV company operating commercial rider-only service in multiple major US cities
Multi-Modal Sensor Fusion	LiDAR + cameras + radar + audio (vs. Tesla's camera-only approach)
HD Map + ML Hybrid	Combines pre-built HD maps with learned perception/planning (vs. purely map-free or purely map-dependent)
Foundation Model Architecture	Production system uses end-to-end trained foundation model, benefiting from Gemini ecosystem
Safety Record	Peer-reviewed studies showing 85%+ crash reduction vs. human drivers
Simulation Depth	Waymo World Model (Genie 3-based) can generate training scenarios from any video input

Comparative Architecture: Waymo vs. Tesla

Dimension	Waymo	Tesla
Sensors	LiDAR + cameras + radar + audio	Cameras only (vision-only)
Maps	Pre-built HD maps (cm-level)	No pre-built maps
Deployment Model	Robotaxi fleet (no private ownership)	Personal vehicle upgrade
Training Data Source	Own fleet (200M+ miles) + simulation (20B+ miles)	Customer fleet (4M+ vehicles, billions of miles)
Hardware Cost Per Vehicle	Higher (LiDAR + sensor suite: est. $10-15K)	Lower (cameras only: ~$400)
Operational Domain	Geo-fenced cities with HD maps	Any road (no geo-fencing)
Autonomy Level	SAE Level 4 (fully driverless in ODD)	SAE Level 2+ (driver supervision required)
Revenue Model	Ride-hailing revenue per trip	Software subscription ($99-199/month FSD)
Compute (Training)	Google TPU pods	Custom Dojo + NVIDIA GPU clusters
Foundation Model	Gemini-derived (EMMA)	Custom transformer (FSD v12+)

Key Advantages Over Competitors

Vs. Cruise (GM): Cruise suspended operations in Oct 2023; Waymo has uninterrupted commercial service
Vs. Tesla: Waymo operates fully driverless (no safety driver); Tesla FSD requires driver supervision
Vs. Baidu Apollo: Waymo has US regulatory relationships and operates in more diverse US cities
Vs. Zoox (Amazon): Waymo has larger fleet, more cities, more autonomous miles
Vs. Mobileye: Waymo operates its own fleet rather than licensing technology to OEMs only

18. Open Source & Datasets

Waymo Open Dataset (WOD)

The Waymo Open Dataset is one of the largest and most diverse autonomous driving datasets publicly available.

License: Non-commercial use

Perception Dataset

Attribute	Detail
Segments	2,030 segments, each 20 seconds at 10 Hz
Total Frames	~390,000
Cameras	5 high-resolution cameras (front, front-left, front-right, side-left, side-right)
LiDAR	Top LiDAR (100-200K points per sweep)
Annotations	3D 7-DOF bounding boxes (vehicles, pedestrians, cyclists, signs) with globally unique tracking IDs
Additional Labels	2D bounding boxes, panoptic segmentation, projected LiDAR on camera, per-pixel camera rays
Label Quality	Independent LiDAR and camera labels (not projections of each other)
Bounding Box Constraints	Zero pitch, zero roll
Calibration	Full intrinsic and extrinsic camera calibration provided

Motion Dataset (WOMD)

Attribute	Detail
Segments	103,354 segments, each 20 seconds at 10 Hz
Duration	570+ hours of unique data
Road Coverage	1,750+ km of roadways
Geographic Coverage	6 US cities
Temporal Windows	9-second windows (1s history + 8s future) with varying overlap
Auto-Labeling	High-accuracy 3D auto-labeling system for all road agents
HD Maps	Corresponding high-definition 3D maps for each scene
Data Format	Sharded TFRecord files containing protocol buffer data
Split	70% training, 15% testing, 15% validation
v1.2.0 Addition	LiDAR points for first 1 second of each 9-second window
v1.3.1 (Oct 2025)	Added `sdc_paths` feature indicating valid future routes for the SDC

End-to-End Driving Dataset (WOD-E2E)

Attribute	Detail
Segments	4,021 segments, each 20 seconds
Duration	~12 hours
Focus	Challenging long-tail scenarios (occurrence frequency <0.03%)
Cameras	8 JPEG cameras (front, front-left, front-right, side-left, side-right, rear, rear-left, rear-right) at 10 Hz
Resolution	~1920x1280 native; typically downsampled to 768x768
Camera FOV	70-90 degrees horizontal per camera; full 360-degree coverage
Includes	High-level routing information, ego states, 360-degree camera views
Evaluation Metric	Rater Feedback Score (RFS) -- measures match to human rater trajectory preference labels
2025 Challenge	WOD-E2E Challenge using held-out test set rater labels

Open Challenges

Waymo hosts annual challenges on its datasets:

Challenge	Dataset	Task
3D Detection	WOD Perception	3D object detection from LiDAR and/or camera
3D Tracking	WOD Perception	Multi-object tracking in 3D
Motion Prediction	WOMD	Multi-agent trajectory forecasting
Sim Agents	WOMD	Simulating realistic agent behavior
Occupancy & Flow Prediction	WOMD	Predicting future occupancy grids and flow fields
End-to-End Driving	WOD-E2E	Vision-based end-to-end driving in long-tail scenarios

GitHub Repositories

Repository	Description
`waymo-research/waymo-open-dataset`	Official dataset tools, tutorials, and evaluation code

Access

Download: waymo.com/open/download
Storage: Google Cloud Storage buckets for programmatic access
API: Python/TensorFlow API with tutorial notebooks
Dataset Format: TFRecord with protocol buffers

SLAM Methods

Methods

Waymo: Comprehensive Autonomous Vehicle Technology Stack ​

Table of Contents ​

1. Company Overview ​

Corporate Identity ​

Funding History ​

Key Milestones ​

2. Vehicle Platform ​

Vehicle Generations ​

Jaguar I-PACE (5th Generation Waymo Driver) ​

Waymo Ojai / Zeekr RT (6th Generation Waymo Driver) ​

Chrysler Pacifica Hybrid (Retired) ​

3. Sensor Suite ​

Generation Comparison ​

LiDAR (In-House Design) ​

Cameras ​

Radar (In-House Imaging Radar) ​

External Audio Receivers (EARs) ​

4. Onboard Compute ​

Compute Architecture ​

Redundancy Philosophy ​

5. Autonomy Software Stack ​

Pipeline Architecture ​

5.1 Perception ​

5.2 Prediction ​

5.3 Planning ​

5.4 Control ​

5.5 End-to-End Trends ​

6. Machine Learning & AI ​

Foundation Models ​

Waymo: Comprehensive Autonomous Vehicle Technology Stack

Table of Contents

1. Company Overview

Corporate Identity

Funding History

Key Milestones

2. Vehicle Platform

Vehicle Generations

Jaguar I-PACE (5th Generation Waymo Driver)

Waymo Ojai / Zeekr RT (6th Generation Waymo Driver)

Chrysler Pacifica Hybrid (Retired)

3. Sensor Suite

Generation Comparison

LiDAR (In-House Design)

Cameras

Radar (In-House Imaging Radar)

External Audio Receivers (EARs)

4. Onboard Compute

Compute Architecture

Redundancy Philosophy

5. Autonomy Software Stack

Pipeline Architecture

5.1 Perception

5.2 Prediction

5.3 Planning

5.4 Control

5.5 End-to-End Trends

6. Machine Learning & AI

Foundation Models