As the automotive industry pushes the boundaries of autonomous driving and ADAS, the integration of AI-driven and deterministic approaches is pivotal in ensuring robust safety and performance. AI has the potential to offer unparalleled advancements in environmental perception, sensor fusion, and the handling of complex scenarios, making it a key enabler for higher levels of automation. However, the inherent unpredictability and lack of traceability in AI decisions necessitate a balanced approach where deterministic methods validate and complement AI outputs. This delicate balance is crucial for developing reliable, safe, and compliant autonomous systems, especially in safety-critical applications.
In this article we speak with Matthias Schulze, Global VP ADAS Product at ECARX GmbH Germany – a European hub based in Böblingen for the Chinese-founded T1, which shares a cofounder with the Geely Group. With additional input from Nexdata and rFpro, we explore the multitude of potential applications for AI within ADAS and AD, and the challenges and opportunities ahead.
Matthias Schulze
Global VP ADAS Product
ECARX
- Environmental Perception: How do you see the balance between AI-driven and deterministic approaches in enhancing environmental perception accuracy, especially in complex driving scenarios like urban environments?
AI driven environment perception provides a major step forward when it comes to assistated or automated driving in complex scenarios such as urban environments. It can be considered as a key enabler. However, as long as it is not possible to track how the AI came to its decision, deterministic approaches are needed to verify the output of the AI model.
Researchers all over the world are working on “traceable AI”, but so far there are no approaches ready to be deployed in commercial products. Therefore, for the time being, we will see a co-existence of AI-driven and deterministic approaches in ADAS and Autonomous Driving systems.
DON’T MISS MATTHIAS’ SESSION AT AUTOSENS EUROPE
BALANCING AI VS DETERMINISTIC APPROACHES FOR ROBUST SAFETY
Matthias Schulze
Global VP ADAS Product
ECARX
On this same topic, Matt -Matt Daley, Technical director, rFpro, says:
AI-driven and deterministic approaches have clear roles in the development of AV systems. AI perception system models (using pattern recognition) are efficient at perceiving and identifying the scene. Physics-based deterministic models are essential for testing and validating AI-based perception algorithms, as they provide a foundation for repeatable and validated results by accurately simulating the physical properties of the environment. The reliability of deterministic models ensures that AI perception systems are tested against a solid, verifiable basis, offering a level of precision and trust that AI generated test data cannot guarantee.
rFpro is working on several projects, including the Centre for Connected and Autonomous Vehicles (CCAV) funded Sim4CAMSens project, to collect real-world data for the purpose of correlating and improving its physics-based simulation platform.
rFpro will be discussing the use of simulation to develop AVs and where it provides the highest value at its roundtable discussion at AutoSens. It will also be offering demonstrations of its AV simulation solution across the event.
CATCH MATT’S SESSION AT AUTOSENS EUROPE
Simulation and Virtualization for Design, Testing, and Validation
Matt Daley,
Technical director, rFpro
rFpro
- Advancing Sensor Fusion: What are the key challenges in advancing sensor fusion technologies, and how can AI and deterministic methods be optimally combined to ensure robust and reliable perception in autonomous vehicles?
AI-driven methods provide new possibilities for low level sensor data fusion and greatly contribute to a better understanding of the vehicle environment and to the decision-making process of the system. But the same applies to the use of AI for sensor data fusion as for environmental perception: As long as we do not know how an AI model comes to its decision, the output of AI needs to be verified by deterministic methods, which leads to a co-existence of both approaches.
Pan Pan, Head of Business Dept, Nexdata went into further detail on this topic:
In the rapidly evolving landscape of autonomous driving, bringing a new ADAS/AD system to market is not only costly but also complex. A crucial step in this process is the validation of autonomous driving functions, which demands the labeling of vast amounts of perception data with the utmost accuracy.
A powerful data annotation platform is key to produce large amounts of annotations in a short amount of time while ensuring a high level of quality. Nexdata, established in 2011, stands at the forefront of AI data solutions, specializing in the provision of superior GT training/validation data for AI research & development in autonomous driving.
Our platform revolutionizes data annotation by blending speed with quality, thanks to our advanced fusion annotation platform capable of handling both dynamic (e.g., cars, pedestrians) and static objects (e.g., lane markings, road space segmentation). We leverage multi-sensor fusion technology to support a variety of data types, including LiDAR, mmWave radar, cameras, and aerial imagery, ensuring comprehensive data alignment and fusion capabilities.
- Generating Edge-Cases: In the context of safety-critical systems, how can AI be used to effectively generate and handle edge cases that deterministic models might miss, and what are the implications for real-world deployment?
We see two applications of AI for generation and handling of edge cases. First, AI can considerably contribute to the identification of edge cases in real life data. Either directly while data is collected or later if the collected data is analyzed and searched for edge cases.
By applying AI-based models, certain requirements for what data will be kept can be defined, such as the number of vehicles and persons or certain driving scenarios. Also, the level of confidence of the AI model can be used as an indicator for edge cases, as Edge cases are cases where AI has difficulty coming to a decision and confidence is therefore lower. Furthermore, AI-based approaches are used to anonymize the collected data.
Recent developments in generative AI open up new possibilities for synthetic Edge Case generation, but one needs to be aware that synthetic data generated by AI might contain errors such as shadows facing in the wrong direction or objects, persons or animals with distorted or amorphous appearances, so that this data needs to be carefully checked before use.
- Automated Parking: What role does the combination of AI and deterministic approaches play in the development of automated parking solutions, and how do these methods ensure precision and safety in tightly constrained spaces?
Advanced parking systems like memory parking or valet parking, where there is no longer a driver in the vehcle while the vehicle searches for a parking space, have much higher requirements on environment perception than systems that are fully supervised by a driver sitting in the car. The vehicle not only has to select a suitable parking spot, it also has to deal with surrounding traffic, people around it, and a multitude of objects like shopping carts, strollers etc.
These systems operate in a much more complex environment than simple automated parking systems have to deal with. As for complex driving situations also for complex parking scenarios. In these complex parking scenarios, as with more complex driving situations, AI-driven environmental perception is required to provide the necessary capability.
- Sensor Calibration and Reliability: How can AI contribute to the continuous calibration and reliability of sensors in autonomous systems, and what deterministic checks are necessary to ensure the accuracy of AI-driven perception?
AI-based sensor calibration tools are on the market for all types of sensors used for ADAS and Autonomous Driving. AI is also the key enabler to novel multi-camera environment perception systems such as wide-baseline-stereo, where the two cameras are no longer solidly connected and camera calibration needs to be constantly adjusted. But there are also research projects ongoing that investigate the use of AI models to detect sensor degradation, for instance through dirt, moisture or fog, or because of a sensor losing its calibration.
Those projects also develop AI-based methods that, in case of sensor degradation or a sensor going out of calibration, initiate suitable countermeasures such as recalibrating a sensor on the fly, sensor cleaning, or a functional degradation of the system, so that safe system operation is still possible.
- Regulatory and Ethical Considerations: With the increasing integration of AI in safety-critical systems, what regulatory and ethical considerations should be taken into account to ensure that both AI and deterministic approaches align with global safety standards?
AI systems have to abide by both existing and new safety requirements and legal structures. This includes frameworks like the EU AI Act amongst others, which guide the development and deployment of AI systems in safety-critical domains. ECARX is also fully aware of the ethical concerns connected to the use of AI in ADAS and Autonomous Driving, and is addressing these together with internal and external experts. This ensures the ethical use of AI especially in safety critical applications.
Very strict regulations apply in Europe and China for data protection and privacy that require the anonymization of all data collected in real life and limit the use and exchange of this data. All ECARX processes in which real life data is collected, used or exchanged are designed so that they strictly follow these regulations to ensure that privacy rights are not violated. The same is expected from ECARX suppliers.
AI systems have to abide by both existing and new safety requirements and legal structures. This includes frameworks like the EU AI Act amongst others, which guide the development and deployment of AI systems in safety-critical domains. ECARX is also fully aware of the ethical concerns connected to the use of AI in ADAS and Autonomous Driving, and is addressing these together with internal and external experts. This ensures the ethical use of AI especially in safety critical applications.
Very strict regulations apply in Europe and China for data protection and privacy that require the anonymization of all data collected in real life and limit the use and exchange of this data. All ECARX processes in which real life data is collected, used or exchanged are designed so that they strictly follow these regulations to ensure that privacy rights are not violated. The same is expected from ECARX suppliers.
- Importance of Collaboration: Why are forums like AutoSens important for the automotive supply chain? What benefits do we glean from coming together to discuss these challenges?
Events like AutoSens are crucial to stay connected with the ADAS and Autonomous Driving community and to learn about the latest developments. They offer a great platform to discuss upcoming challenges with experts from all over the world. They also provide visibility for our company in an environment that is constantly changing.
Balancing AI and deterministic approaches is not just a technical necessity, but a strategic imperative for the future of autonomous driving. Whilst AI can be leveraged to drive innovation, deterministic methods provide the essential safety net needed to meet stringent regulatory and ethical standards. As the industry moves forward, the co-existence of these approaches will continue to be essential in ensuring the safety, reliability, and public trust in autonomous systems. By fostering collaboration and continuous innovation, the automotive sector can harness the full potential of AI while upholding the rigorous standards required for real-world deployment.
To hear more from experts on the above topics, and to participate in the networking and collaboration opportunities helping to advance industry connectivity and progress, join us at AutoSens Europe.
Connect on this topic in the exhibition
AutoSens and InCabin include a technology exhibition with an array of technical demonstrations, vehicle demonstrations, buck demos from world leading companies. Engineers who come to AutoSens and InCabin have the opportunity to not only discuss the latest trends but actually get their hands on the tech and see it in action.
