Chapter 4: Simulation, HD Maps, and Regulations

Section 9: Simulation Technologies

Testing autonomous vehicles exclusively on public roads is impractical and dangerous. Billions of miles of driving data are needed to statistically validate safety claims, but accumulating that mileage physically would take decades and cost lives. Simulation addresses this gap by providing safe, scalable, and repeatable virtual environments in which autonomous driving (AD) stacks can be developed, trained, and validated far more efficiently than on-road testing alone.

This section covers the leading simulation platforms, digital twin methodologies, and synthetic data generation techniques that form the backbone of modern AD development pipelines.

9.1 CARLA: Open-Source AD Simulator

CARLA (Car Learning to Act) is the most widely adopted open-source simulator purpose-built for autonomous driving research. Originally released by Intel Labs, Toyota Research Institute, and the Computer Vision Center (CVC) Barcelona, CARLA provides a flexible environment for development, training, and validation of AD systems.

CARLA v0.9.16 (2025) Key Features

The 2025 release of CARLA v0.9.16 introduced several transformative capabilities that bring the platform closer to production-grade simulation:

CARLA's open-source nature and extensible Python/C++ API have made it the de facto standard for academic AD research, with over 3,000 published papers referencing the platform.

9.2 NVIDIA DRIVE Sim

While CARLA serves the open-source community, NVIDIA DRIVE Sim represents the industry-grade simulation platform built on the Omniverse Blueprint architecture. DRIVE Sim provides physically accurate sensor simulation spanning camera, radar, and LiDAR rendering with ray-traced fidelity.

Core Capabilities

• DRIVE Map: Provides access to road network digital twins hosted on Omniverse. These map assets can be converted into fully driveable simulation environments, ensuring that virtual testing scenarios accurately reflect real-world road geometry, signage, and lane markings.
• Virtual Reconstruction: One of two AI-based approaches to scene creation. Real-world driving scenarios are reconstructed as complete synthetic 3D scenes, enabling deterministic replay with controllable variations (weather, lighting, traffic density).
• Neural Reconstruction: The second AI-based approach, which extends real sensor data through neural simulation. Rather than building scenes from scratch, this method augments captured data with neural rendering to generate novel viewpoints and conditions from existing recordings.

Industry Adoption

Major automotive OEMs have integrated DRIVE Sim into their development workflows:

• Jaguar Land Rover (JLR): Built their entire 2025 vehicle lineup on an NVIDIA co-developed software-defined architecture. All testing is conducted in simulation before on-road validation, significantly reducing development cycles.
• Mercedes-Benz: Runs thousands of virtual driving scenarios through DRIVE Sim to validate their DRIVE PILOT L3 system, covering edge cases that would be impractical or dangerous to test physically.

9.3 Digital Twins

A digital twin in the AD context is a faithful reproduction of a real-world road environment in virtual space. Unlike generic simulation environments that use procedurally generated or artist-created scenes, digital twins aim for precise geometric and visual correspondence with specific physical locations.

Key Characteristics

• Geometric Fidelity: Centimeter-accurate 3D reconstructions of roads, intersections, buildings, and infrastructure derived from survey-grade LiDAR scans and photogrammetry.
• Neural Rendering: Modern digital twins employ neural radiance fields (NeRF) and 3D Gaussian Splatting for high-fidelity scene reconstruction from real sensor data. These techniques generate photorealistic novel views, enabling new scenario creation from limited real-world captures.
• Dynamic Elements: Beyond static geometry, digital twins increasingly incorporate realistic traffic flow models, pedestrian behavior patterns, and environmental variations (time-of-day, weather, seasonal changes).

Digital twins enable safe and efficient testing by providing environments where rare and dangerous scenarios (e.g., pedestrian dashing onto a highway, multi-vehicle pile-ups, adverse weather on specific roads) can be tested repeatedly without physical risk.

9.4 Synthetic Data Generation

Synthetic data generation extends simulation beyond testing into the training data pipeline. Rather than relying exclusively on expensive, manually labeled real-world datasets, teams generate precisely labeled synthetic data at scale.

NVIDIA DRIVE Sim for Synthetic Data

NVIDIA DRIVE Sim generates synthetic datasets specifically designed to measure deep neural network (DNN) sensitivity to new sensor positions. When an OEM modifies sensor placement on a new vehicle model, DRIVE Sim can generate training data reflecting the new viewpoints without physical data collection campaigns.

Novel View Synthesis Research

Emerging research in novel view synthesis enables teams to emulate different camera configurations from existing driving data. By incorporating these synthetically generated viewpoints into training data, perception systems can be made more robust to sensor placement variations, improving generalization performance across vehicle platforms.

This approach represents a significant cost reduction over traditional data collection methods, where each new sensor configuration would require months of physical driving and annotation effort.

Simulation-to-Deployment Pipeline

The following diagram illustrates the end-to-end pipeline from simulation through validation to real-world deployment:

This pipeline reflects how modern AD development has shifted from a predominantly physical testing paradigm to a simulation-first approach, where the vast majority of scenario coverage is achieved virtually before any on-road validation occurs.

Section 10: HD Maps / High-Definition Maps

High-definition maps have been a cornerstone of traditional autonomous driving systems, providing centimeter-accurate road geometry, lane boundaries, traffic sign positions, and semantic information that supplements real-time sensor perception. However, the economics and limitations of HD maps have driven a fundamental rethinking of how AD systems relate to map data.

Challenges of Traditional HD Maps

Despite their precision, HD maps present several critical challenges that limit their scalability:

Prior to 2022, the AD industry focused primarily on HD map geometric accuracy, with traditional Advanced Driver-Assistance Systems (ADAS) relying on preset rules tied to map features. This approach, while effective in well-mapped corridors, could not scale globally.

Mapless Approach

The mapless approach to autonomous driving represents one of the most significant architectural shifts in the industry. Rather than depending on pre-built map databases, mapless systems construct their understanding of the driving environment in real time from onboard sensors alone.

Pioneers and Adopters

Tesla's Full Self-Driving (FSD) system pioneered the mapless approach, demonstrating that an end-to-end neural network could navigate complex urban environments without relying on HD map priors. This approach has since been adopted by multiple major Chinese manufacturers:

• Xpeng: Deployed mapless navigation across Chinese cities.
• Huawei AITO: Integrated mapless AD in their intelligent driving system.
• GAC Aion: Adopted mapless capabilities for their EV lineup.
• Li Auto: Transitioned from HD-map-dependent to mapless highway and city driving.

Evolution Trajectory

The industry transition follows a clear progression:

1. Map Prior Knowledge (pre-2022): AD systems heavily dependent on pre-built HD maps for localization, planning, and semantic understanding.
2. Real-Time Map Construction (2022-2024): Systems began constructing local maps on-the-fly from sensor data, reducing (but not eliminating) dependence on stored maps.
3. World Models (2024-present): The latest paradigm where neural networks build internal world representations that go beyond geometric mapping to include predictive understanding of scene dynamics, physics, and agent behavior.

This evolution mirrors a broader shift in ADAS algorithm design from rule-driven approaches (if lane boundary detected, then steer) to data-driven approaches (learn optimal behavior from millions of driving examples).

Enabling Technologies

Advanced rendering and reconstruction techniques are transforming what maps can represent:

• 3D Gaussian Splatting: Enables real-time, high-quality 3D scene reconstruction from sparse viewpoints, evolving maps from "past records" to "future previews" that can predict how a scene will look from new positions.
• Neural Radiance Fields (NeRF): Provides dense, photorealistic scene representations from limited camera inputs, enabling both map construction and simulation scenario generation from the same underlying technology.

Hybrid Mapping

Rather than choosing exclusively between HD maps and mapless approaches, many companies are pursuing hybrid strategies that combine the strengths of both paradigms.

Hybrid Strategy Components

• HD Map Precision: Retained for well-mapped corridors (highways, major urban arteries) where centimeter-accurate lane geometry adds measurable safety value.
• SD+ Map Broad Accessibility: Standard-definition maps enhanced with key semantic features (SD+ maps) provide cost-efficient coverage for areas where full HD mapping is economically infeasible.
• Online/Dynamic Mapping: Bird's-eye-view (BEV) representations generated in real time from onboard sensors create local maps that complement or override stored map data. These BEV maps are generated through neural networks that fuse multi-camera inputs into a unified overhead perspective.

Industry Perspective

NavInfo, one of China's largest mapping companies, has proposed "inheriting map spatiotemporal cognition into World Models." This vision represents the convergence of traditional mapping expertise with neural world modeling: rather than discarding decades of cartographic knowledge, the industry is encoding that understanding into the training data and architectures of next-generation AD systems.

Section 11: Regulations and Safety Standards

The deployment of autonomous vehicles on public roads requires not only technological maturity but also comprehensive regulatory frameworks that define responsibilities, safety requirements, and operational boundaries. As of 2025, regulatory approaches vary significantly across major markets, reflecting different legal traditions, risk tolerances, and industrial policy priorities.

11.1 Japan

Japan has positioned itself as a global leader in autonomous driving regulation through proactive legislative reform and government-backed demonstration programs.

Legislative Framework

In April 2023, Japan enacted an amended Road Traffic Act that legalized Level 4 "Specified Automated Driving." This made Japan one of the first countries to establish a comprehensive legal framework for vehicles operating without any human driver in designated areas and conditions.

Government Deployment Targets

RoAD to the L4 Project

The "RoAD to the L4" project, associated with Japan's Strategic Innovation Promotion Program (SIP), began in September 2021. The project coordinates model region demonstrations to build technical capability and social acceptance for autonomous driving services.

Key achievements to date include:

• Eiheiji Town: Became the first location in Japan to operate L4 autonomous vehicles on public roads, using low-speed electric carts for last-mile transportation in a rural community.
• Hitachi BRT: Bus Rapid Transit integration with autonomous driving technology on dedicated bus lanes.
• Haneda Airport Area: Autonomous shuttle services connecting airport terminals with surrounding facilities.
• Shin-Tomei Expressway: Truck platooning demonstrations on Japan's newest expressway, testing L4 truck convoys for long-haul logistics.

11.2 United States

The United States takes a decentralized, state-led approach to autonomous driving regulation, resulting in a patchwork of rules that varies significantly by jurisdiction.

State-Level Regulation

As of 2025, 38+ states have enacted some form of AV legislation. The regulatory landscape varies widely:

• Some states issue commercial fleet permits allowing robotaxi operations (e.g., California, Arizona, Texas).
• Others require remote driver availability as a condition for driverless operation.
• A small number maintain effective bans on fully driverless vehicles on public roads.

Federal Regulation (NHTSA)

The National Highway Traffic Safety Administration (NHTSA) has gradually adapted federal safety standards to accommodate autonomous vehicles:

• March 2022: Revised safety requirements to allow vehicles without traditional driving controls (steering wheels, pedals) to comply with Federal Motor Vehicle Safety Standards (FMVSS). This removed a major regulatory barrier for purpose-built autonomous vehicles.
• June 2025: Expanded FMVSS exemptions, allowing deployment of safety-validated AVs that do not meet traditional design requirements (e.g., vehicles designed without a driver seat or manual controls), provided they demonstrate equivalent safety through alternative means.
• April 2025: Relaxed L2 crash reporting requirements, narrowing mandatory reporting to incidents involving death, hospitalization, pedestrian contact, or airbag deployment. This reduced the administrative burden on L2 system operators while maintaining coverage of serious incidents.

11.3 European Union

The European Union has pursued a harmonized regulatory approach through the United Nations Economic Commission for Europe (UNECE) framework, supplemented by member-state initiatives.

UN Regulation No. 157 (UN-R157)

Effective from January 2023, UN-R157 is the first internationally binding regulation for automated driving. Applied in both Japan and the EU, it permits:

• Level 3 automated driving on highways at speeds up to 130 km/h.
• Lane change maneuvers during automated driving mode.

UN-R157 establishes requirements for system activation/deactivation, transition demands to the driver, minimum risk maneuvers, and data recording obligations.

Member State Initiatives

• Germany: Legalized L4 autonomous driving in 2021, becoming the European leader in AV regulation. In 2025, Germany introduced mandatory liability insurance specifically for L4+ vehicles, clarifying the insurance framework for commercial autonomous services.
• France: In 2025, France mandated black box recorders (Event Data Recorders) for all autonomous vehicles, ensuring that comprehensive data is available for post-incident investigation and liability determination.

Rules for L4/L5 unmanned vehicles are currently in active development at the UNECE level, with a focus on defining safety validation methodologies and operational design domain requirements.

11.4 China

China has adopted an aggressive, centrally coordinated approach to autonomous driving deployment, combining national policy targets with city-level pilot programs.

National Targets

China's target is for 30%+ of new car sales to feature L3+ autonomous driving capabilities by 2025, reflecting the scale of ambition in the Chinese market.

Regulatory Framework

In November 2023, China launched a national pilot program that clarified commercialization pathways for L2 through L4 autonomous driving. A distinctive feature of the Chinese approach is the emphasis on vehicle-infrastructure-cloud seamless integration, leveraging China's 5G infrastructure and smart city investments to support cooperative autonomous driving.

Pilot Zone Cities

20+ cities have been designated as pilot zones for autonomous driving testing and deployment, including Beijing, Shanghai, and Shenzhen. These zones feature dedicated testing roads, regulatory sandboxes, and streamlined permitting processes.

Industry Progress

BYD obtained the first conditional L3 test license in China in July 2023. By June 2025, nine companies including NIO, Changan, and GAC had prepared for L3 road testing, reflecting rapid industry mobilization in response to regulatory clarity.

11.5 Safety Standards

Beyond national regulations, international safety standards provide the technical framework for ensuring autonomous vehicles operate safely. Three standards form the core of AD safety assurance:

Complementary Safety Framework

ISO 26262 and SOTIF together form a comprehensive safety assurance framework:

• ISO 26262 covers hazards caused by system failures (e.g., a sensor stops working, a processor crashes, a communication link drops).
• SOTIF (ISO 21448) covers hazards caused by performance limitations (e.g., a camera is blinded by direct sunlight, a radar misclassifies a guardrail, a pedestrian detection model fails on unusual clothing).

Together, these standards address the full spectrum of safety concerns: what happens when the system breaks, and what happens when the system works as designed but the world presents scenarios beyond its validated capability envelope.

RSS complements both by providing a formal, mathematically verifiable definition of safe behavior that can be used as both a planning constraint and a post-hoc evaluation metric.

Chapter Summary

This chapter covered three interconnected pillars of the autonomous driving ecosystem that extend beyond the core perception-planning-control software stack:

Key Takeaways

• Simulation has become the primary development and validation methodology for AD systems. CARLA provides open-source accessibility while NVIDIA DRIVE Sim delivers production-grade fidelity. Digital twins and synthetic data generation close the sim-to-real gap and reduce data collection costs.
• HD Maps are transitioning from a mandatory prerequisite to one component in a spectrum of mapping strategies. The mapless approach pioneered by Tesla has shifted the industry toward real-time perception and world models, while hybrid strategies combine the best of both approaches.
• Regulations vary significantly by region. Japan and Germany lead in L4 legal frameworks, the US takes a state-led approach, the EU pursues harmonization through UNECE, and China combines national targets with city-level pilot zones. ISO 26262, SOTIF, and RSS together define the safety engineering framework.

Looking Ahead

In Chapter 5, we will examine the remaining challenges facing the autonomous driving industry, the market outlook, and the broader societal implications of this transformative technology.