探索全新的 mydSPACE 服务: 许可证管理、全新的 dSPACE 帮助以及用于用户文档的 dSPACE 聊天机器人。请访问我们的专用支持页面,获取更多信息和支持。 链接了解更多并获取支持

Validation Requires New Testing Strategies

The first part of this blog article explained how the validation of ADAS/AD functions presents new challenges for vehicle manufacturers and suppliers. The proportion of software is increasing, AI is gaining importance, a growing number of heterogeneous sensors is becoming tightly integrated, and E/E architectures are taking on a more central role. As a result, the effort required to reliably and cost-effectively validate safety-critical functions is also increasing.

The second part of this blog article focuses on how to tackle these tasks in practice using dSPACE tools. A significant shift in testing methodology toward virtual validation, also known as shift-left testing, is key. Another important aspect is digital homologation. In a holistic ADAS/AD test strategy, HIL serves as a platform that bridges the gap between virtual development and real vehicle testing. It creates reproducible conditions for the robust validation of safety-critical functions deployed on real ECUs and HPCs.
 

What an ADAS/AD HIL Must Be Able to Do

As with traditional automotive domains, an ADAS/AD system must perceive the test environment as realistic. Therefore, the system under test (SUT) must not detect any discrepancy between laboratory conditions and actual vehicle operation. The structure of the ADAS/AD software stack is crucial here. Modular architectures place different demands on the HIL than monolithic end-to-end AI stacks do (Figure 1).
With modular stacks, individual functions – such as perception or planning – can be validated separately. In end-to-end AI stacks, perception, decision-making, and control are much more closely intertwined. The test environment must provide consistent sensor data and also enable a closed-loop control system. This is precisely where data replay and closed-loop HIL differ most significantly. 
 

Figure 1: Modular and end-to-end architectures place different demands on ADAS/AD validation. 

Data Replay: Close to the Reality of the Recorded Drive

Modular ADAS/AD stacks currently dominate up to Level 2 autonomy. In this context, initial testing often focuses on the performance of perception, i.e., object detection and sensor data fusion. These functions can be largely validated independently of downstream modules. Since AI-based perception models are predominantly trained using real-world data, they are sensitive to variations in sensor data, such as noise, lighting conditions, textures like asphalt or vegetation, and backgrounds.
For the SUT to accept the HIL environment, the system must be fed real sensor data that is as unaltered as possible. This includes raw camera data, as well as radar, lidar, and ultrasonic data. This data is usually collected during test drives and other campaigns, is time-stamped, and must be reproduced with exact temporal correlation during replay. Only then can sensor fusion generate a consistent picture of the environment at all times. 
 

Figure 2: Data replay delivers recorded sensor and bus data to the SUT with precise timing.    

Why Time Synchronization Is Key to Quality

The dSPACE data replay solution (Figure 2) bridges the gap between real-world testing and virtual testing on HIL particularly effectively. A carefully coordinated combination of software, real-time components, Environment Sensor Interface Units (ESI Units), and SCALEXIO hardware ensures precise time synchronization and very low latency. The sensor input can be synchronized via an internal real-time clock or an external time master using the Precision Time Protocol (PTP).
However, the quality of the replay depends heavily on the input data. In-vehicle loggers may introduce drift or jitter. Non-deterministic communication protocols such as SOME/IP also make an exact 1:1 replay difficult, especially during a HIL component test when other ECUs are not available as real communication partners. dSPACE compensates for such effects with time corrections and tailored simulation models for service-based protocols. This ensures that the replay remains deterministic, accurate, and precise.
 

Get More Value from Recorded Data

Recording driving data in the field is expensive and time-consuming. That is why it should remain usable for as long as possible throughout the development project. dSPACE supports this through data conversion and targeted manipulation during playback. This enables existing data sets to be adapted to software updates, architectural changes, and new safety requirements without having to re-record every situation in the vehicle.

Data replay is particularly well-suited for realistic testing of perception functions and modular ADAS/AD stacks. It provides reproducible conditions, eliminates the real-to-sim gap, and makes valuable field data available in the lab. However, this approach reaches its limits as soon as the SUT is required to actively intervene in the scene.
  
 

Closed-Loop HIL: When the System Interacts with the Environment

Data replay has limitations when testing components of the downstream stack, such as prediction, planning, and control. In open-loop replay, the SUT cannot influence the environment. If an end-to-end (E2E) model makes a driving decision that differs slightly from the driver’s decision in the original recording, it expects a new perspective. However, the recorded sensor data does not reflect this new perspective. Consequently, the model and the scene diverge, inducing a distribution shift for downstream model parts. Added to this is the risk of causal confusion. E2E models learn statistical correlations and can take undesirable shortcuts. For example, a model might associate braking actions with brake lights rather than correctly identifying the obstacles. Without feedback from a responsive environment, such effects are only partially visible in an open-loop system. Comprehensive validation therefore requires a closed-loop control system. A further critical point at which using data replay alone is insufficient is in so-called long-tail scenarios. Another limitation of using data replay alone is its inability to validate long-tail scenarios. These uncommon edge cases, such as unusual traffic behaviors or rare environmental conditions, are critical for ADAS/AD validation. While data replay is limited to previously observed events, closed-loop simulation enables the systematic generation and evaluation of scenarios that may never appear in collected data.

Core Requirements for Closed-Loop HIL

  • Realistic and physically plausible sensor simulation
  • Real-time-capable simulation and sensor data generation for stable closed-loop operation
  • Precise temporal correlation of sensor data
  • Absolute synchronization between all HIL components and the SUT

Figure 3: In a closed-loop system, the vehicle model, environment, sensor simulation, interfaces, and SUT interact in real time.

How dSPACE Closes the Control Loop

In the dSPACE ADAS/AD HIL solution (Figure 3), the individual components are precisely matched to one another. dSPACE’s Automotive Simulation Models (ASM) provide vehicle dynamics and environmental simulation as ground truth and a single source of truth. The models run in real time on the SCALEXIO platform. AURELION uses this information, supplements it with material properties and high-definition 3D scenes, and generates realistic data for camera, radar, lidar, and ultrasonic sensors on GPUs using physics-based sensor models.
To ensure that simulation and sensor data generation remain fast enough, AURELION, the high-performance GPUs, and the PC platform are optimally parametrized according to individual project needs. Camera images, radar echoes, and lidar point clouds generated by the sensor simulation PC are available at its output. These are then efficiently transmitted to the SUT via ESI Units and SCALEXIO, just as in data replay. The goal is to minimize the sim-to-real gap while maintaining stable, closed-loop real-time operation. One key enabler for achieving this is the synchronization, which is more challenging in closed-loop mode than in replay mode. There are no recorded time stamps on which the data feed can rely. Instead, each system component aligns with a common time base. ASM transfers simulation data to AURELION at fixed time intervals. AURELION generates the corresponding data for all sensors in parallel and holds it until it can be fed into the SUT synchronously. In addition, the SUT can request adjustments to the cycle times via a control channel connected to an ESI unit. Consequently, the simulation can be adapted accordingly, and all HIL components are resynchronized to the updated timing.

The dSPACE HIL solution fulfills the stringent requirements of closed-loop ADAS/AD testing. This enables reliable validation of AI stacks for automated driving on target hardware, regardless of whether the underlying architecture is modular or end-to-end.
 

Fault Injection for Robust ADAS/AD Testing   

HIL testing is particularly valuable when errors can be injected in a targeted and reproducible manner. dSPACE provides comprehensive fault-injection capabilities including: 

  • Electrical and communication faults 
  • Sensor-level disturbances
  • Deterministic or random pixel and line noise
  • Frame drops, frame tearing, timing errors, and jitter
  • Errors in embedded data and statistical data, such as frame counters, histograms, and CRC
  • Frame cropping, subsampling, and letterboxing
  • Typical wire and coaxial cable faults, such as open circuits, short circuits, and stuck-at-GND/VBAT conditions
  • Communication faults on vehicle networks and buses

Figure 4: Simulated error input helps test sensor and interface behavior under controlled conditions.

Turnkey, Modular, and Efficient Project Implementation

A comprehensive ADAS/AD-HIL setup can integrate components from multiple vendors. dSPACE modular design and open interfaces fully support this approach. In practice, however, this often increases the need for coordination. Having many project partners or frequently switching vendors can extend development times and increase project risks.
To address these challenges dSPACE offers complete turnkey systems for ADAS/AD-HIL and data replay. In this case, dSPACE acts as the central point of contact, ensuring seamless interaction between hardware, software, sensors, and interfaces. This reduces technical complexity, shortens commissioning times, and simplifies the resolution of open issues with partners.
Another advantage is the flexibility of the platform. Shared hardware and standardized software make switching between data replay and ADAS/AD-HIL easy. Depending on the project’s requirements, the system can be designed for both use cases or adapted later. This ensures that the investment remains usable, even if the test strategy, software architecture, or validation focus changes. 
Incidentally, customer requirements often overlap. In many cases, this allows us to leverage ADAS/AD HIL systems that are already largely built and equipped with standardized components. Based on these pre-engineered platforms, only the customer-specific elements need to be implemented, which can significantly accelerate solution delivery and commissioning.
 

Figure 5: Turnkey ADAS/AD-HIL systems combine hardware, software, partner integration, and global support.

Next Step: AI, DevOps, and Standards

The functions for autonomous driving continue to evolve toward end-to-end ADAS/AD stacks. Due to their more natural, driver-like behavior, these stacks are considered an important path forward. At the same time, the need for explainable artificial intelligence (XAI) is growing. XAI can reduce the “black box” nature of today’s end-to-end AI, make model decisions more transparent, and support the validation of safety-critical functions. dSPACE recognizes the growing importance of closed-loop HIL simulation for validation of E2E AI and XAI. To address this, dSPACE is further reducing the sim-to-real gap and enhancing simulation model credibility.
This includes approaches such as neural reconstruction, in which recorded log data is converted into an editable, three-dimensional, world using Neural Radiance Field (NeRF) models or the advanced Gaussian Splatting algorithm, both of which are AI-based. If an E2E model deviates from the recorded driving path during simulation, photorealistic real-time camera perspectives can be generated. The simulation thus responds more closely to the logic of the system under test.
Meanwhile, dSPACE is integrating into DevOps pipelines via CI/CD processes and supporting industry standards such as OpenDRIVE and OpenSCENARIO.
Automated testing plays a key role in CI/CD workflows. With the Python-based Test Automation SDK, dSPACE provides unified access to VEOS and SCALEXIO. In addition to traditional testing tools such as Pytest and Robot Framework, LLM-generated tests and future agentic AI systems can also be integrated. These are expected to support the entire testing lifecycle, from requirements analysis to test generation, execution, and evaluation. 
 

Figure 6: Trends such as neural reconstruction, simulation credibility, TA-SDK, CI/CT, and OpenX standards increase the efficiency of validation. 

Conclusion: The Right Testing Approach for Every Development Phase

Data replay and closed-loop HIL have different objectives but complement each other perfectly. Data replay makes real-world field data reproducible, making it particularly well-suited for perception and sensor data fusion tests. Closed-loop HIL builds on this approach by including interaction, real-time sensor simulation, and dynamic system behavior. This enables the validation of even complex planning, control, and end-to-end AI functions.
For development projects, the right combination is what matters most. The earlier and more consistently tests are conducted virtually, the more effectively risks can be mitigated, test coverage increased, and costs controlled. dSPACE supports this approach with scalable solutions that integrate real ECUs, HPCs, simulation, sensor data, and automation into a consistent validation strategy.
 

About the Author

Gregor Hordys

Gregor Hordys

Business Field Manager HIL Testing, dSPACE

通过 dSPACE时事通讯服务了解最新信息。

通过 dSPACE 时事通讯服务,我们将随时向您通报当前的使用案例、新解决方案和产品以及培训和活动。在此注册,免费订阅。

Enable form call

At this point, an input form from Click Dimensions is integrated. This enables us to process your newsletter subscription. The form is currently hidden due to your privacy settings for our website.

External input form

By activating the input form, you consent to personal data being transmitted to Click Dimensions within the EU, in the USA, Canada or Australia. More on this in our privacy policy.