March 9, 2018 (JST)

National Research and Development Agency
Japan Aerospace Exploration Agency (JAXA)

JAXA started collaboration with Japan Automobile Manufacturers Association (JAMA) on the quality assurance of artificial intelligence (AI) computer designed to drive autonomous automobiles.
In the course of development of spacecraft and other pursuits, JAXA has acquired and built on the scientific knowledge of the software quality assurance, including Independent Verification and Validation (IV&V)^*1 and hazard analysis process^*2. To make further applications of the expertise, in the period through March 2019, JAXA partners with JAMA in administrative and procedural activities implemented in the quality of autonomous system of AI platform to ensure the requirements and goals for the product. The challenge of self-driving system is taking outputs combined with predicted state of the surroundings to decide the right trajectory for the vehicle. A full measure of safety assurance is needed to enable total autonomy in driving.

Through the collaborative initiative, JAXA offers its insight on the safety analysis technology that coordinates the human-machine interaction. This helps establish a framework of quality assurance of AI computer used to drive cars and perform other tasks. JAXA also expects to provide its quality strategy proved by IV&V to guarantee the safe and sound operations of the software and the overall system.

By working with Japanese automobile industry, JAXA aims to obtain the knowledge necessary to develop safe AI software onboard spacecraft. JAXA also wishes to obtain useful knowledge to standardize the safety evaluation process for the rapidly expanding AI system across many other fields and industries.

*1 IV&V: measures taken to assure the safe, reliable operation of mission critical software. Identification and countermeasure of errors are done, independent of system developers.

*2 Embodied in Systems-Theoretic Accident Model and Process (STAMP) and STAMP based Process Analysis (STPA). Hazard analysis model and process based on the interactivity of control and surrounding circumstances.