Sensor fusion is an essential prerequisite for self-driving cars, and one of the most critical areas in the autonomous vehicle (AV) domain. Now it has become clear that a common agreement or standard can help to reduce liability risks and the risk of wrong development. OEMs, Tier 1 and Tier 2 suppliers could also benefit from more efficient collaboration. This is what renowned industry experts discussed at the latest virtual Chapter Event by The Autonomous and co-host BASELABS.
“Sensor fusion flaws can jeopardize the safety of a self-driving vehicle’s entire system,” said Chairman Ricky Hudi after The Autonomous’ Chapter Event. “The state of sensor data fusion does not yet provide a reliable basis for safe AV – more has to be done. Industry-wide collaboration could help lift advanced driver assistance systems and later autonomous driving to the necessary safety level.”
“We must make automated vehicles safe under all conditions,” says Robin Schubert, Co-founder and Managing Director of BASELABS. “Therefore, we support The Autonomous: we believe in cooperation in the industry, want to exchange best practices and experiences and work together to design reference solutions that later can also be accepted by standards. We see a clear need for an agreement on how we can prove that sensor data can be reliable.”
Standardization of data fusion was a key discussion point among the experts who also took questions from the audience.
Cornelius Bürkle, Research Scientist at Intel Corporation, was convinced that safety should not be part of competition. From his perspective, industry and organizations should agree on which models in system architectures are needed and what degree of safety they should cover. This could help decide on the trade-off between the availability of sensor data and safety.
Carlo van Driesten, Systems Architect for Virtual Test & Validation at BMW, explained during the discussion that standardization should not prevail over innovation. But open standards could be helpful to inspire and support innovation. In his talk, he presented the ASAM Open Simulation Interface (OSI) published by BMW as a GitHub open source project as a best practice example. The challenge now is to further develop the Open Scenario Standard.
According to Alexander Scheel, Sensor Fusion Engineer for Automated Driving at Bosch, there is consensus in different research communities when it comes to KPIs, for example. But there is little or no exchange across disciplines such as classical tracking and the growing machine learning community. Bert Auerbach, Vice President at FDTech, mentioned that in the area of testing, requirements are often not fully known to suppliers. He suggested that the Operational Design Domain (ODD) for the model must be defined specifically. Ronny Cohen, CEO of LeddarTech, said because there is no SAE Level 3/4 system yet, it might be too early to talk about standards for it, but the industry is coming closer to this and is probably near to standards for SAE Level 2 systems.
Marcus Obst, Head of Business Development at BASELABS, presented possible benefits for OEMs and Tier 1 and Tier 2 suppliers of a standardized data fusion architecture. Standards could make supplier offers comparable for OEMs and help to formalize the offering process. On the other hand, suppliers would need much less tailored engineering of existing components for the input and output interfaces. A once standardized data fusion architecture under ISO 262626 and ISO/PAS 21448 would eliminate the need for further standardization by multiple stakeholders. As already applicable for SAE Level 2 and 2+, this could be a first step to build on for higher SAE levels.
Read more about the further discussion on data fusion at this Chapter Event on The Autonomous website.