Nuro: Autonomous Vehicle Technology Stack — Exhaustive Technical Writeup

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
Safety Architecture
Fleet Operations
Key Partnerships
Regulatory
Pivot / Restructuring
Research & Publications
Competitive Position

1. Company Overview

Founding & Leadership

Detail	Value
Founded	September 2016
Founders	Dave Ferguson & Jiajun Zhu
HQ	1300 Terra Bella Avenue, Mountain View, CA
Employees	~1,009 (January 2026)
Peak Headcount	~1,400 (late 2022)
Total Funding	~$2.34 billion across 7 rounds
Current Valuation	$6 billion (April 2025)
Peak Valuation	$8.6 billion (2021)

Dave Ferguson — Co-founder and Co-CEO. Holds an MS and PhD in Robotics from Carnegie Mellon University and a BS in Computer Science and Mathematics from the University of Otago (New Zealand). Joined Google's self-driving car project (now Waymo) in 2011 as Principal Machine Learning Engineer. Led the planning group for CMU's winning team in the 2007 DARPA Urban Grand Challenge. Holds over 100 patents and has published over 60 academic papers with 24,600+ citations. One of his planning algorithms is currently used for long-range autonomy on NASA's Mars Rovers.

Jiajun Zhu — Co-founder and Co-CEO. Holds an MS in Computer Science from the University of Virginia and a BS from Fudan University. Was a founding team member of Google's self-driving car project, serving as Principal Software Engineer from 2008 to 2016. Led perception technology development and built/led Waymo's simulation efforts. Holds over 100 patents.

Funding History

Round	Date	Amount	Lead / Key Investors	Post-Money Valuation
Series A	June 2017	$92M	Greylock, Banyan; NetEase founder Ding Lei	—
Series B	February 2019	$940M	SoftBank Vision Fund	$2.7B
Series C	November 2020	$500M	T. Rowe Price	$5.0B
Series D	November 2021	$600M	Tiger Global, Google, Kroger	$8.6B
Series E (initial)	April 2025	$106M	Kindred Ventures and others	~$6.0B
Series E (close)	August 2025	$97M + NVIDIA/Uber investment totaling $203M	Uber, NVIDIA	$6.0B

Notable investors include: SoftBank Vision Fund, T. Rowe Price, Tiger Global, Fidelity, Greylock Partners, Google, Kroger, NVIDIA, and Uber.

Key Milestones

Year	Milestone
2016	Company founded by ex-Waymo engineers
2018 (Jan)	First product launch: electric self-driving delivery vehicle concept with $92M funding
2018 (Jun)	First partnership announced (Kroger)
2018 (Aug)	First pilot: autonomous grocery delivery in Scottsdale, AZ (Toyota Prius + R1)
2019 (Feb)	$940M Series B from SoftBank
2020 (Feb)	First-ever NHTSA autonomous vehicle exemption granted (R2)
2020 (Feb)	R2 begins testing in Houston, TX
2021 (Aug)	$40M investment in southern Nevada manufacturing and test facilities
2022 (Jan)	R3 unveiled with external airbag; BYD manufacturing partnership announced
2022 (Sep)	10-year autonomous delivery deal with Uber Eats
2022–2023	Multiple rounds of layoffs (~30% workforce reduction)
2024 (Sep)	Business pivot: from delivery bot manufacturer to autonomy technology licensor
2025 (Mar)	Lenovo partnership for autonomous driving on NVIDIA DRIVE
2025 (Jul)	Nuro-Uber-Lucid robotaxi partnership announced (20,000+ vehicles)
2025 (Aug)	$203M Series E close; NVIDIA and Uber invest
2026 (Jan)	Lucid-Nuro-Uber robotaxi unveiled at CES 2026; autonomous on-road testing begins

2. Vehicle Platform

Nuro pioneered the purpose-built, zero-occupant autonomous delivery vehicle — a form factor with no passenger compartment, approximately half the width of a standard sedan, designed exclusively for goods transport.

Generation Comparison

Specification	R1 (Gen 1)	R2 (Gen 2)	R3 (Gen 3)
Year Introduced	2018	2020	2022
Status	Retired prototype	Deployed	Paused (pivot)
Length	~6 ft (1.8 m)	108 in (274 cm)	~20% smaller than avg passenger car
Width	~Half a sedan	43 in (110 cm)	Similar to R2
Height	~6 ft (1.8 m)	73 in (186 cm)	—
Weight	~1,500 lb (680 kg)	2,535 lb (1,150 kg)	—
Max Payload	~12 grocery bags	419 lb (190 kg)	500 lb (227 kg) / 24 bags
Cargo Volume	Baseline	Baseline	2x R2
Max Speed	Low-speed	25 mph (40 km/h)	45 mph (72 km/h)
Battery	—	31 kWh (custom; all-day operation)	All-day single charge
Drivetrain	Electric	Electric	Electric (BYD Blade Battery)
External Airbag	No	No	Yes (Autoliv)
Thermal Compartments	No	Limited	Yes (heating/cooling)
Manufacturing	In-house	In-house	BYD (Lancaster, CA) + Nuro (NV)
NHTSA Exemption	No	Yes (first ever)	Planned but paused

R1 — First Generation

The R1 was Nuro's proof-of-concept prototype, demonstrating the world's first goods-only autonomous vehicle on public roads. It weighed approximately 1,500 lb, stood just over 6 feet tall, and was about half the width of a sedan. It had capacity for roughly 12 grocery bags. It was first deployed alongside modified Toyota Prius vehicles in the 2018 Kroger/Fry's pilot in Scottsdale, Arizona.

R2 — Second Generation

The R2 was Nuro's first production-intent vehicle and the first autonomous vehicle to receive a federal exemption from NHTSA. Key design features:

Zero-occupant design: No steering wheel, pedals, mirrors, or windshield wipers
Compact footprint: 108 in long x 43 in wide x 73 in high — narrower than a standard vehicle lane
Dual compartments: Two reconfigurable cargo bays with heating/cooling
Sensor array: 12 high-definition cameras (360-degree overlapping FOV), roof-mounted LiDAR, and 14 additional LiDAR/radar units
Custom 31 kWh battery: Nearly double the size of the R1 battery, enabling all-day operation
Redundant systems: Dual computing, braking, steering, power, and sensor systems
SAE Level 4 autonomy: Fully driverless within its operational design domain
Energy-absorbing front panel: Purpose-designed crumple structure for pedestrian protection

R3 — Third Generation

Unveiled in January 2022, the R3 was designed to be the first automotive-grade delivery vehicle produced at scale. Key advances:

Doubled cargo volume vs R2 with modular storage inserts
500 lb payload capacity with temperature-controlled compartments
45 mph top speed (up from R2's 25 mph), enabling operation on more road types
External pedestrian airbag by Autoliv — inflates across the entire front of the vehicle upon imminent collision detection
Multi-modal sensor suite including cameras, radar, LiDAR, and thermal cameras with redundant 360-degree coverage
BYD manufacturing partnership: BYD provides the Blade Battery, electric motors, electronic controls, and displays; assembles the hardware platform at its Lancaster, CA plant; Nuro completes autonomy system integration at its Nevada facility

R3 production was paused following the September 2024 business pivot.

Post-Pivot: Robotaxi Platform (Lucid Gravity)

Following the 2024 pivot, Nuro's technology is being integrated into passenger vehicles. The flagship program is the Nuro-Uber-Lucid robotaxi:

Base vehicle: Lucid Gravity SUV
Seating: Up to 6 passengers with generous luggage space
Sensor integration: High-resolution cameras, LiDAR, and radar with 360-degree awareness; sensors integrated into the vehicle body and a low-profile roof-mounted "Halo" module
Compute: NVIDIA DRIVE AGX Thor
Autonomy: Nuro Driver Level 4 self-driving system
Fleet commitment: 20,000+ vehicles across dozens of US and international markets
Timeline: On-road testing began early 2026; first rides via Uber app expected late 2026

3. Sensor Suite

Nuro employs a multi-modal sensing architecture designed for maximum redundancy and pedestrian safety. The suite has evolved through vehicle generations, with the latest next-generation architecture moving from traditional rotary LiDAR to distributed solid-state sensors.

Sensor Modalities

Sensor Type	Role	Key Characteristics
Solid-State LiDAR	Precise 3D depth measurement, object detection	Small form factor; no moving parts; enhanced long-range detection; improved side and rear visibility vs rotary LiDAR
Automotive Cameras	Object classification, semantic understanding	High-definition; 360-degree overlapping coverage (R2: 12 cameras); differentiate vehicles, pedestrians, cyclists, etc.
Imaging Radar	Distance/velocity measurement, inclement weather operation	Long-range; elevation data; improved velocity and angular resolution; detect small/distant objects in rain, fog, and snow
Thermal Camera	Pedestrian detection via infrared (heat)	Detects body heat instead of visible light; particularly effective at night and for partially occluded pedestrians; front-facing
Microphone	Emergency vehicle siren detection	Enhanced audio processing for directional siren identification
IMU	Inertial navigation, vehicle state estimation	Upgraded to ASIL-D rated (highest automotive safety integrity level)

R2 Sensor Configuration

12 high-definition cameras with 360-degree overlapping field of view
1 roof-mounted primary LiDAR
14 additional LiDAR/radar sensors for object detection and velocity estimation
Thermal camera (front-facing)

Next-Generation Sensor Architecture (Nuro Driver Platform)

The next-generation architecture was designed for cross-platform deployment (delivery bots, robotaxis, passenger vehicles):

Distributed solid-state LiDAR: Replaces single rotary unit; multiple small sensors mounted around the vehicle perimeter for enhanced coverage and reliability; fewer moving parts for better manufacturing and durability
Automotive-grade cameras: Arranged for redundant 360-degree FOV with overlapping coverage zones
Long-range imaging radar: High-resolution sensors for detecting distance and depth of small, distant objects including road debris, animals, and vulnerable road users
Thermal cameras: Independent pedestrian detection modality
Halo module (robotaxi): Low-profile roof-mounted sensor pod integrating cameras, LiDAR, and radar into the vehicle body design

The unified perception system ingests synchronized data from approximately 30 individual sensors, including multiple long-range and short-range cameras, LiDARs, and radars.

All sensors demonstrated significant improvements in challenging weather conditions and long-distance detection of road debris, animals, and vulnerable road users (cyclists, pedestrians) during validation testing.

4. Onboard Compute

Current Platform: NVIDIA DRIVE AGX Thor

In March 2024, Nuro announced that the next generation of its Nuro Driver system would be built on the NVIDIA DRIVE Thor compute platform.

Specification	Detail
SoC	NVIDIA DRIVE Thor
Architecture	NVIDIA Blackwell (designed for transformer, LLM, and generative AI workloads)
AI Compute (single SoC)	Up to 1,000 TFLOPS (FP8/INT8)
AI Compute (dual SoC)	Up to 2,000 TFLOPS (FP8/INT8)
Operating System	NVIDIA DriveOS (safety-certified)
Functional Safety	Designed for highest levels of automotive functional safety; 15,000+ engineering years invested by NVIDIA across the full stack
Consolidation	Sensor processing, AI-first autonomy, and safety-critical components unified on a single centralized compute unit

Architecture Design Rationale

By building on DRIVE Thor, Nuro consolidates all intelligent vehicle functions — sensor processing, AI inference, planning, and safety monitoring — into a centralized SoC. This eliminates the need for multiple discrete compute modules, reducing system complexity, weight, power consumption, and cost.

The Blackwell architecture is purpose-designed for the large transformer and foundation model workloads that form the core of Nuro's AI-first autonomy approach. The platform supports:

Real-time multi-sensor fusion across ~30 sensors
Unified perception foundation model inference
Prediction and planning model execution
Safety monitoring and fallback systems
Vehicle control and actuation

Lenovo AD1 Domain Controller

Through the Lenovo partnership (March 2025), Nuro's autonomy software is also being integrated with Lenovo's AD1 domain controller, which is itself built on the NVIDIA DRIVE AGX platform featuring the DRIVE AGX Thor SoC and DriveOS. This provides an additional hardware pathway for OEM customers.

FTL Model Compiler Framework

Nuro developed the Faster Than Light (FTL) Compiler Framework, a custom ML model optimization toolchain for deploying models to onboard compute:

Ingests models from multiple training frameworks (TensorFlow, PyTorch)
Applies multiple industry compilers in a single compilation pass within a customizable Python environment
Supports multi-GPU inference using pipeline parallelism (splits model via subgraphs across devices)
Supports quantization for reduced precision inference
Has delivered significant reductions in CPU compute utilization, GPU compute utilization, and GPU memory consumption
Provides a unified optimization platform — a single code change delivers performance improvements to all models
Enables non-ML-infra teams (e.g., onboard performance) to optimize CPU utilization and upgrade CUDA APIs

Redundancy

All vehicle platforms feature redundant computing, braking, steering, power, and sensor systems to ensure the vehicle can safely come to a controlled stop if any single system fails.

5. Autonomy Software Stack

The core autonomy platform is called Nuro Driver — a complete Level 4 self-driving system that integrates perception, prediction, planning, and control. It has been successfully integrated into seven different vehicle platforms including delivery robots, passenger vehicles, and trucks, and has been proven on roads across three US states.

Stack Components

┌─────────────────────────────────────────────────────┐
│                   NURO DRIVER                       │
├──────────┬──────────┬───────────┬──────────────────┤
│ Percep-  │ Predic-  │ Planning  │ Vehicle          │
│ tion     │ tion     │           │ Control          │
├──────────┴──────────┴───────────┴──────────────────┤
│          Unified Perception Foundation Model        │
├────────────────────────────────────────────────────-┤
│          Sensor Fusion (Voxel Feature Space)        │
├──────────┬──────────┬───────────┬──────────────────┤
│ Camera   │ LiDAR    │ Radar     │ Thermal /        │
│ Encoder  │ Encoder  │ Encoder   │ Other Encoders   │
├──────────┴──────────┴───────────┴──────────────────┤
│          HD Map & Localization Layer                │
├─────────────────────────────────────────────────────┤
│          NVIDIA DRIVE Thor / DriveOS               │
└─────────────────────────────────────────────────────┘

Perception

Nuro's perception system uses an ML-first architecture with robust fallbacks. It fuses data from cameras, LiDAR, radar, and thermal sensors to detect and track objects including pedestrians, vehicles, cyclists, animals, and road debris.

Key characteristics:

Unified Perception Model: A single foundation model processes all sensor inputs rather than separate per-sensor pipelines
Independent multimodal sensor encoders feed into a unified voxel feature space
Sensor fusion module transforms features from native formats into the shared voxel representation
Temporal modeling aligns spatial features across time steps (T to T-n) and generates spatial-temporal features with stateful temporal consistency
Input from ~30 individual sensors (multiple long-range and short-range cameras, LiDARs, radars)
Vision Transformer (ViT) architecture integration (boosted emergency vehicle detection by 5%)

Prediction

The prediction module analyzes the intents of other traffic agents (vehicles, pedestrians, cyclists) and forecasts their future behavior using machine learning models. This is critical for the delivery use case where the vehicle frequently operates in residential neighborhoods with less predictable pedestrian behavior (children, pets, etc.).

Planning

The planning system determines the vehicle's path and actions based on perceived environment state and predicted agent behaviors. Key features:

Reinforcement learning-assisted plan selection: Improved pullover and pullout maneuvers for smoother operation
Human feedback reward models: Incorporated into behavior models with active learning to encourage safer driving in various conditions
Delivery-specific priorities: Optimized for frequent stops, residential navigation, curbside approaches, and double-parking avoidance — different from passenger AV planning which prioritizes route efficiency and passenger comfort
Zero-occupant advantage: The vehicle can maneuver and brake in ways that would be uncomfortable or unsafe for passengers (e.g., harder emergency braking), reducing collision risk

Control

Low-level vehicle control translating planned trajectories into actuator commands (steering, acceleration, braking). The zero-occupant form factor allows more aggressive safety maneuvers since there is no concern about passenger comfort or injury from abrupt movements.

AI-First Architecture

Nuro describes its approach as "AI-first," meaning:

One to two very large foundational AI models perform many tasks (mapping, localization, perception, prediction, planning) in a single integrated architecture
This contrasts with traditional AV stacks that use many separate, hand-tuned modules
The approach enables rapid adaptation to new environments, vehicle types, and functions
Significantly reduces deployment timelines for new OEM customers

6. Machine Learning & AI

Unified Perception Foundation Model

Nuro's core ML architecture centers on a Unified Perception Foundation Model — a single large model that processes all sensor modalities and outputs multiple perception tasks simultaneously.

Architecture details:

Independent Multimodal Sensor Encoders: Each sensor modality (camera, LiDAR, radar) has its own encoder that processes raw sensor data into intermediate features
Image Encoder Pretraining: The image encoder is pretrained on large-scale image datasets, improving performance on all downstream tasks and facilitating integration with other foundation models
Sensor Fusion Module: Transforms encoded features from native sensor formats into a unified voxel feature space, generating multi-modal spatial features
Temporal Module: Aligns spatial features from time T to T-n and fuses them into spatial-temporal features; performs stateful temporal modeling for consistency, conditioning on spatial-temporal features at T and task features/queries from T-1 to T-n
Task Heads: Downstream task-specific heads for detection, tracking, classification, segmentation, etc.

Design goals:

Cross-platform compatibility: A single foundational model can be trained once and exported/deployed across different vehicle platforms with efficient post-training adaptation
Open vocabulary capabilities being integrated using feature extractors from multimodal language models (VLMs)
End goal: a fully end-to-end learnable autonomy system meeting L4 performance and safety requirements

Reinforcement Learning

Nuro has built a scalable RL training infrastructure for autonomous driving:

Distributed training: Scales across GPU/TPU resources and parallel simulations
Training throughput: Can train models on decades worth of driving experience in hours
Abstracted communication and simulation: Researchers can plug in new agent architectures and algorithms to test state-of-the-art RL methods
Closed-loop training: RL agents train in simulation with realistic dynamics and traffic
Applications: Plan selection (pullover/pullout), behavior modeling, driving policy refinement
Human feedback: Reward models incorporate human preferences via active learning for safer driving behavior

Training Infrastructure

Component	Detail
Accelerators	NVIDIA V100, A100, H100 GPUs; Google TPUs
Preferred for transformers	TPUs (more cost-effective for XLA-compatible transformer models)
Hosting	Google Kubernetes Engine (GKE)
Scheduling	Custom Nuro ML Scheduler
Training type	Large-scale distributed training; many jobs per day

Nuro ML Scheduler features:

Accepts training jobs from users and allocates appropriate accelerator resources
Complements (does not replace) the Kubernetes scheduler
Intelligently preempts jobs to yield to higher-priority work and enforce group quotas
Automatically selects the best accelerator type per job based on quota and resource availability
Handles multi-accelerator job placement across heterogeneous hardware

Model Deployment Pipeline

Training on cloud TPU/GPU clusters via Nuro ML Scheduler
Compilation through FTL Model Compiler Framework (supports TensorFlow, PyTorch)
Optimization including quantization, multi-GPU pipeline parallelism
Deployment to onboard NVIDIA DRIVE Thor compute
Validation through simulation and closed-course testing before road deployment

7. Mapping & Localization

HD Mapping Pipeline

Nuro builds its own high-definition maps for all operational areas through an automated cloud-based pipeline:

Aspect	Detail
Generation method	Automated pipeline combining robotics, ML, and human verification
Resolution	Centimeter-level detail
Content	Lanes, traffic lights, crosswalks, road features, curbs, driveways
Data sources	On-road sensor data (collected via Prius sensor stacks and AV fleet), aerial imagery, satellite data
Scalability	Designed to support rapid expansion to new delivery areas and cities

Scalable HD Mapping Approach

Nuro introduced a novel approach that fuses offline map priors with online sensor data:

The model consumes both out-of-date offline semantic map features and real-time online sensor measurements
It learns to pass through the offline HD map prior when correct but remains robust to map changes and low-quality labeling
This addresses the fundamental challenge of HD map maintenance — maps become stale as road infrastructure changes

Geospatial Foundation Model

Nuro is developing a geospatial foundation model for mapping and localization:

Combines real-time perception data with low-cost map priors
Enables:
- Precision global localization (centimeter-level position tracking at all times)
- Online map feature inference (real-time map updates from sensor data)
- Real-time sensor calibration

Learned Localization

The localization architecture features:

Online encoder: Consumes onboard LiDAR point cloud data
Geospatial encoder: Consumes aerial digital surface models (DSM) and/or satellite imagery
Cross-correlation alignment: Embedding images are aligned by computing cross-correlation over a search window of possible x, y, and theta offsets
This "aerial-ground" localization approach bridges satellite/aerial map data with street-level sensor data, reducing dependence on expensive HD map maintenance

8. Simulation Platform

Nuro employs a three-tier testing methodology: virtual simulation, closed-course physical testing, and real-world on-road validation.

Virtual Simulation

Capability	Detail
Data source	Previously collected real-world sensor data from public roads
Environment	Virtual reconstruction of real environments using sensor data
Augmentation	Synthetic augmentation of real data to enhance realism and scenario coverage
Smart agents	AI-driven traffic participants with realistic behavior
Fault injection	Ability to inject faults and measure latency for "what if" scenarios
Metrics	2,000+ user-configurable autonomy metrics for evaluation and analytics
Scale	Millions of test scenarios generated automatically

Foretellix Partnership

In January 2024, Nuro partnered with Foretellix (now part of NVIDIA) for advanced virtual testing:

Automatically generates millions of relevant and meaningful test scenarios
Ensures proper Operational Design Domain (ODD) coverage
Uncovers edge cases that might pose safety risks
Reduces R&D costs through more efficient virtual validation
Leverages Foretellix's scenario-based testing standards

Closed-Course Testing Facility

Nuro invested $40M in facilities in southern Nevada, including a world-class closed-course track:

Facility Detail	Specification
Location	Las Vegas Motor Speedway, NV
Size	74 acres
Scenario types	Pedestrian avoidance, pet avoidance, cyclist interaction, shared roadway navigation
Additional testing	Environmental tests, vehicle systems validation
Purpose	Bridge between simulation and real-world deployment

Real-World Validation

L4 driverless releases undergo testing across 40 validation categories before public road approval
Strict control and risk assessment procedures govern all deployments
Safety Committee conducts daily performance and safety monitoring
Crash reporting within 24 hours (NHTSA requirement)

Testing Pipeline

Simulation (millions of scenarios)
    → Closed-Course (controlled physical tests)
        → Limited Real-World (geofenced ODD)
            → Expanded Deployment (validated ODD)

9. Cloud & Data Infrastructure

Cloud Platform

Nuro's cloud infrastructure runs primarily on Google Cloud Platform (GCP):

Component	Technology
Compute orchestration	Google Kubernetes Engine (GKE)
ML training accelerators	NVIDIA V100, A100, H100 GPUs; Google Cloud TPUs
ML job scheduling	Custom Nuro ML Scheduler (on top of Kubernetes)
Simulation	Cloud-based replay and scenario generation
HD map pipeline	Automated cloud pipeline (robotics + ML + human verification)
Data storage	Large-scale sensor data lakes (camera, LiDAR, radar logs from fleet)
Evaluation tooling	2,000+ configurable metrics computed at scale from on-road and simulation logs

Data Pipeline

Collection: Fleet vehicles collect multi-sensor data (camera, LiDAR, radar, IMU, GPS) during all on-road operations
Upload: Data transferred to cloud storage infrastructure
Processing: Automated pipelines process raw sensor data for training, mapping, and evaluation
Training: ML models trained on cloud GPU/TPU clusters via Nuro ML Scheduler
Evaluation: Simulation re-plays real-world scenarios with new model versions; 2,000+ metrics computed
Deployment: Validated models compiled via FTL and pushed to vehicle fleet

Resource Management

The Nuro ML Scheduler provides sophisticated resource management:

Multi-tenant scheduling across heterogeneous accelerator types
Priority-based preemption for critical training jobs
Quota enforcement across engineering groups
Automatic accelerator type selection based on model architecture and availability
Support for distributed training across multiple nodes

10. Safety Architecture

Nuro's safety philosophy is fundamentally shaped by its zero-occupant vehicle design — eliminating the most vulnerable person in any vehicle collision (the passenger) allows radical re-optimization of the entire safety envelope toward protecting people outside the vehicle.

Zero-Occupant Safety Advantages

Advantage	Explanation
No passenger risk	Zero occupants means no risk of passenger injury in any scenario
Aggressive braking	Vehicle can execute emergency stops that would injure passengers in a conventional car
Sacrificial maneuvers	Vehicle can prioritize external safety even at cost of vehicle damage
Lower mass	Lighter vehicle = lower kinetic energy in collision
Lower speed	Purpose-built for 25-45 mph operation (R2/R3)
No human error	Eliminates distracted/impaired/fatigued driving

External Pedestrian Airbag (R3)

The R3 features an industry-first exterior pedestrian airbag developed in partnership with Autoliv (global automotive safety supplier):

Airbag covers the entire front surface of the vehicle when inflated
Triggers upon detection of imminent frontal collision
Optimized to reduce the force of impact on pedestrians and cyclists
Supported by Nuro's network of self-cleaning sensors providing 360-degree awareness
Complements the energy-absorbing front panel crumple structure

Energy-Absorbing Front Panel

All Nuro vehicles feature a purpose-designed front structure that absorbs collision energy, reducing force transmitted to a struck pedestrian. Unlike passenger vehicles where the front structure must also protect occupants, Nuro's entire front can be optimized exclusively for external impact absorption.

System Redundancy

System	Redundancy
Computing	Dual compute systems
Braking	Redundant braking
Steering	Redundant steering
Power	Redundant power supply
Sensors	Overlapping sensor coverage from multiple modalities

If any primary system fails, redundant systems ensure the vehicle can execute a safe stop.

Safety Testing & Validation

40 validation categories tested before every L4 driverless release
Safety Committee conducts daily performance monitoring
Partnership with Foretellix for automated edge-case scenario discovery
Closed-course testing at 74-acre Las Vegas facility for physical scenario validation
24-hour crash reporting to NHTSA
Regular meetings with NHTSA during exemption periods
Community outreach in deployment areas

IMU Safety Rating

The next-generation sensor architecture includes an upgraded inertial measurement unit rated at ASIL-D (Automotive Safety Integrity Level D) — the highest safety integrity level defined in ISO 26262 for automotive functional safety.

11. Fleet Operations

Deployment History

Location	Period	Status	Partners	Notes
Scottsdale, AZ	Aug 2018 – 2019	Closed	Kroger/Fry's	First pilot; Toyota Prius + R1 vehicles
Phoenix, AZ	2018 – Oct 2022	Closed	Kroger	Depot closed Oct 2022; removed from commercial roadmap
Houston, TX	Apr 2019 – present	Active	Kroger, Walmart, Domino's, FedEx, Uber Eats	Primary operational hub; R2 deployment; 70% increase in deployment linear miles
Mountain View / Palo Alto, CA	2020 – present	Active	Uber Eats	HQ area operations; 83% increase in deployment linear miles
San Francisco Bay Area	2026 – present	Testing	Uber/Lucid robotaxi	Robotaxi prototype testing with safety operators

Operational Capabilities

Driverless operation: Fully autonomous Level 4 delivery service without safety drivers in approved areas
Delivery workflow: Customer orders online, vehicle navigates to store, store staff loads cargo, vehicle delivers to customer address, customer retrieves from compartment via app
Temperature control: Heated and cooled compartments keep food fresh
All-day operation: Custom battery enables full-day operation on a single charge
Tens of thousands of deliveries completed for partners including Kroger and Uber Eats

Expansion Post-Pivot

Following the 2024 pivot, Nuro's fleet operations are transitioning:

Delivery operations: Continue in Houston and Mountain View/Palo Alto
Robotaxi testing: Engineering prototypes with safety operators in San Francisco Bay Area (early 2026)
Future robotaxi launch: First Uber-native robotaxi rides expected in a major US city via the Uber app in late 2026, with plans for dozens of US and international markets

12. Key Partnerships

Retail & Delivery Partners

Partner	Year	Relationship	Details
Kroger	2018	Strategic grocery delivery	First commercial partner; pilot at Fry's in Scottsdale, AZ; expanded to Houston; Kroger also a Series D investor
Domino's	2019	Pizza delivery	Pilot in Houston; autonomous R2 pizza delivery
Walmart	2019	Grocery delivery	Pilot in Houston using R2 and Toyota Prius vehicles; customer opt-in program
CVS Pharmacy	2020	Prescription delivery	Autonomous prescription delivery pilot
FedEx	2021	Parcel logistics	Pilot in Houston; Nuro's first foray into package/parcel delivery
Uber Eats	2022	Food delivery (10-year deal)	Autonomous delivery service in California and Texas

Technology & Manufacturing Partners

Partner	Year	Relationship	Details
BYD	2022	R3 manufacturing	BYD provides Blade Battery, motors, electronics; assembles at Lancaster, CA plant; Nuro completes autonomy integration in NV. Partnership paused in 2024 pivot.
NVIDIA	2024	Compute platform & investor	DRIVE Thor SoC for Nuro Driver; NVIDIA invested in Series E; 15,000+ engineering years of safety development in DRIVE platform
Autoliv	2022	External airbag	Designed and supplies exterior pedestrian airbag for R3
Foretellix	2024	Simulation & testing	Automated scenario generation for virtual validation; ODD coverage assurance
Lenovo	2025	Domain controller & go-to-market	AD1 domain controller on NVIDIA DRIVE AGX; accelerating commercial AV deployment for OEMs

Mobility & Robotaxi Partners

Partner	Year	Relationship	Details
Uber	2022 / 2025	Delivery + robotaxi	10-year delivery deal (2022); $300M investment in Lucid for robotaxi program (2025); committed to 20,000+ robotaxis; investor in Nuro Series E
Lucid Motors	2025	Robotaxi vehicle platform	Lucid Gravity SUV as base platform; Uber investing $300M in Lucid; unveiled at CES 2026; testing began early 2026

Serve Robotics Clarification

Nuro and Serve Robotics (NASDAQ: SERV) are separate, competing companies in the autonomous delivery space. Serve Robotics originated as Postmates X (Postmates' robotics division, founded 2017), was acquired by Uber with Postmates in 2020, and spun out as an independent company in 2021. Serve focuses on smaller sidewalk robots operating at low speeds within a 2-mile radius, while Nuro operates larger road vehicles at higher speeds. Both companies have relationships with Uber but are independent competitors.

13. Regulatory

NHTSA Exemption (February 2020) — First Ever

Nuro's R2 became the first autonomous vehicle in history to receive a federal exemption from the National Highway Traffic Safety Administration (NHTSA).

Detail	Specification
Vehicle	R2 (R2X designation in filing)
SAE Level	Level 4 (highly automated)
Max speed	25 mph
Drivetrain	Electric
Occupants	Zero (no human occupant by design)
Exempted standard	FMVSS No. 500 (low-speed vehicle requirements)
Exemption basis	(1) Facilitates development of low-emission vehicle without unreasonably lowering safety; (2) Compliance would prevent sale of a vehicle with equal or better overall safety
Production limit	No more than 5,000 R2 vehicles during 2-year exemption period
Reporting requirements	24-hour crash reporting; regular NHTSA meetings; community outreach in operational areas
Significance	First time NHTSA granted exemption for an autonomous vehicle without standard safety/control equipment (steering wheel, pedals, mirrors, etc.)

What Was Exempted

The R2 lacks conventional vehicle controls because it has no human occupant:

No steering wheel or control column
No brake/accelerator pedals
No side mirrors
No windshield wipers
No standard occupant protection (airbags, seatbelts)

These absences would normally violate Federal Motor Vehicle Safety Standards, but NHTSA determined the overall safety level was at least equal to conventional vehicles given the zero-occupant design and comprehensive AV safety systems.

California DMV Permits

Permit	Significance
Driverless testing permit	One of only 6 companies approved for driverless testing in California
Autonomous vehicle deployment permit	First-ever CA DMV AV deployment permit for commercial delivery service
Commercial operations	One of only 3 companies with permits allowing commercial autonomous operations in CA (alongside Mercedes-Benz and Waymo)

Texas Operations

Texas has a more permissive regulatory framework for autonomous vehicles, allowing Nuro to operate with fewer permit restrictions. Nuro has conducted commercial deployments and driverless testing in Houston and surrounding areas.

14. Pivot / Restructuring

Timeline of Changes

Date	Event
Late 2022	First round of layoffs; headcount begins declining from ~1,400 peak
2023	Multiple additional layoff rounds; ~30% (340 employees) cut in largest single reduction
2024 (early)	R3 production with BYD paused
September 2024	Formal business pivot announced
2025–2026	New partnerships (NVIDIA, Uber, Lucid, Lenovo) under new strategy

From Builder to Licensor

Before the pivot (2016–2024):

Nuro designed, manufactured, and operated its own custom delivery robots
End-to-end vertical integration: vehicle design, sensor suite, autonomy software, fleet operations
Revenue model: delivery service fees from retail partners

After the pivot (September 2024–present):

Nuro licenses its Nuro Driver autonomous driving software and sensor platform to OEMs, suppliers, and mobility providers
No longer manufactures its own vehicles
Paused BYD partnership for R3 production
Workforce stabilized at ~1,000 employees (down from 1,400 peak)

Two Go-to-Market Strategies

Full L4 Autonomous Driving Product: Complete self-driving system for goods delivery and passenger mobility services (e.g., Uber robotaxi program)
OEM/Supplier Licensing: Working with automakers and Tier 1 suppliers to build automated driving products for consumer vehicles ranging from Level 2 to Level 4

Rationale

The pivot reflects the reality that building and operating a custom delivery vehicle fleet is extremely capital-intensive. By licensing technology, Nuro can:

Reach massive scale through OEM production volumes (20,000+ Lucid robotaxis vs hundreds of R2s)
Reduce capex requirements (no manufacturing)
Access new markets (passenger vehicles, trucks, consumer cars)
Generate recurring software licensing revenue
Leverage the Nuro Driver's proven cross-platform adaptability (integrated into 7 vehicle platforms)

15. Research & Publications

Founders' Academic Background

Dave Ferguson:

PhD and MS in Robotics, Carnegie Mellon University
BS in Computer Science and Mathematics, University of Otago
60+ peer-reviewed academic publications
100+ patents
24,600+ citations on Google Scholar
Led CMU's planning group for the winning 2007 DARPA Urban Grand Challenge entry
Algorithm currently used for Mars Rover long-range autonomy (NASA)
Research areas: motion planning, path planning, field robotics, machine learning for robotics

Jiajun Zhu:

MS in Computer Science, University of Virginia
BS from Fudan University
100+ patents
Founding team member of Google's self-driving car project (Waymo)
Research areas: perception systems, simulation for autonomous vehicles

Published Technical Blog Posts

Nuro's engineering team has published detailed technical content on their approaches:

Topic	Key Content
The Nuro Autonomy Stack	End-to-end description of perception, prediction, planning, control
Unified Perception Model	Foundation model architecture with voxel representations and multimodal sensor encoders
Scaling ML Training at Nuro	Nuro ML Scheduler, TPU/GPU infrastructure, distributed training
FTL Model Compiler Framework	Custom ML inference optimization, multi-GPU pipeline parallelism, quantization
Exploring HD Mapping that Scales	Offline-online map fusion, scalable HD map generation
Learned Localization	Aerial-ground localization bridging satellite imagery with LiDAR
Enabling Reinforcement Learning at Scale	Distributed RL training, closed-loop simulation, decades of experience in hours
Safety @ Nuro: Our Vehicles	Redundancy design, pedestrian safety philosophy
Safety @ Nuro: Testing	Three-tier testing methodology (simulation, closed-course, real-world)
Next-Generation Sensor Architecture	Solid-state LiDAR, ASIL-D IMU, distributed sensor design

Patent Portfolio

Both co-founders hold 100+ patents each (200+ combined), covering:

Autonomous vehicle perception and detection
Motion planning and trajectory optimization
Simulation and virtual testing
Sensor fusion architectures
Delivery vehicle design and safety systems
HD mapping and localization

16. Competitive Position

Autonomous Delivery Landscape

Company	Vehicle Type	Operating Surface	Speed	Status (as of early 2026)
Nuro	Half-width road vehicle / Robotaxi platform	Public roads	25–45 mph	Active (pivoted to licensing); robotaxi testing underway
Starship Technologies	Small 6-wheeled robot	Sidewalks	~4 mph	Active; 2,000+ bots deployed; 8M+ deliveries completed globally
Amazon Scout	Small 6-wheeled robot	Sidewalks	Low	Discontinued (2022); failed to meet performance targets
Serve Robotics (SERV)	Small sidewalk robot	Sidewalks	Low	Active; publicly traded (NASDAQ: SERV); Uber Eats partnership
Kiwibot	Small sidewalk robot	Sidewalks	Low	Active; university campus focus
Waymo (via partners)	Full-size passenger vehicles	Public roads	Full speed	Active in ride-hailing; exploring delivery
Gatik	Box trucks	Public roads	Full speed	Active; middle-mile B2B delivery

Nuro's Competitive Advantages

First-mover on NHTSA exemption: Only company to receive federal autonomous vehicle exemption; regulatory credibility with NHTSA and California DMV
Purpose-built zero-occupant design: Enables unique safety optimizations (external airbag, aggressive braking, pedestrian-first design) impossible in passenger vehicles
Road operation (not sidewalk): Operates at vehicle speeds on public roads, enabling faster delivery over longer distances than sidewalk bots
L4 technology platform: Proven autonomy system integrated across 7 vehicle platforms; licensable to OEMs
Deep partnerships: 10-year Uber Eats deal; 20,000+ vehicle Uber/Lucid robotaxi commitment; NVIDIA strategic investment; Lenovo go-to-market collaboration
Founder pedigree: Both co-founders are senior Waymo alumni with deep AV expertise; combined 200+ patents
Pivot to platform play: Licensing model allows scale through OEM manufacturing rather than capital-intensive self-manufacturing

Nuro's Challenges

Valuation decline: Down from $8.6B peak (2021) to $6B (2025) — a 30% reduction
Workforce reduction: From ~1,400 to ~1,000 employees through multiple layoff rounds
R3 not in production: Third-generation delivery vehicle paused; BYD partnership suspended
Limited delivery fleet: R2 fleet remains small compared to Starship's 2,000+ deployed bots
Pivot risk: Transitioning from vehicle builder to technology licensor is a fundamental business model change
Competition from Waymo: Waymo's scale, Google backing, and proven robotaxi operations represent a major competitive threat in the licensing/platform space
Revenue generation: The company has raised $2.34B but has not demonstrated sustained revenue at scale

Strategic Position

Nuro has transformed from a purpose-built delivery robot company into an autonomous driving technology platform. The Uber-Lucid robotaxi program (20,000+ vehicles, late 2026 launch) represents the most significant test of this strategy. If successful, it would make Nuro one of the largest autonomous driving technology suppliers globally, alongside Waymo and Cruise's technology stacks.

The company's core technical differentiators — the AI-first unified perception model, the cross-platform Nuro Driver system, and the NVIDIA DRIVE Thor integration — position it as a viable autonomy supplier for OEMs seeking to add L2-L4 capabilities without developing the technology in-house.

SLAM Methods

Methods

Nuro: Autonomous Vehicle Technology Stack — Exhaustive Technical Writeup ​

Table of Contents ​

1. Company Overview ​

Founding & Leadership ​

Funding History ​

Key Milestones ​

2. Vehicle Platform ​

Generation Comparison ​

R1 — First Generation ​

R2 — Second Generation ​

R3 — Third Generation ​

Post-Pivot: Robotaxi Platform (Lucid Gravity) ​

3. Sensor Suite ​

Sensor Modalities ​

R2 Sensor Configuration ​

Next-Generation Sensor Architecture (Nuro Driver Platform) ​

4. Onboard Compute ​

Current Platform: NVIDIA DRIVE AGX Thor ​

Architecture Design Rationale ​

Lenovo AD1 Domain Controller ​

FTL Model Compiler Framework ​

Redundancy ​

5. Autonomy Software Stack ​

Stack Components ​

Perception ​

Prediction ​

Planning ​

Control ​

AI-First Architecture ​