Yrvann, autonomous systems research engineer at Safran Tech

From automobiles to aerospace

Science and research have held a fascination for Yrvann for as long as he can remember. So, it was only natural that he enroll in an engineering school He specialized in systems, imagery and signals engineering and artificial intelligence, working notably on speech emotion recognition during a first research internship with PSA (now Stellantis). During his final year in engineering school, he pursued a master’s degree in automation and robotics research for critical real-time embedded systems. He then completed a PhD on the architecture of a survival system for safe and agile autonomous vehicles, under a CIFRE industrial research training agreement.

Yrvann joined Safran Tech’s Signals and Information Technology research group in February 2018 to work on design, research and technology demonstrations at the Paris-Saclay facility. The associated autonomous vehicles laboratory matures concepts for the development of highly autonomous systems. “We want to show their potential for land vehicles and then use the technology building blocks for aeronautical applications,” says Yrvann.

 

Artificial intelligence in autonomous systems

“An autonomous driving system is a complex and critical real-time system that’s not the same thing as the driver-assistance systems in today’s vehicles, no matter how sophisticated,” Yrvann explains. “It must incorporate all tactical and operational features needed to drive a vehicle, so that it can at least partially replace a human driver for long periods. In particular, it must monitor and understand the driving environment so it can respond appropriately. Developing such a system taps into a range of disciplines, including signal processing, automation, robotics and artificial intelligence.”

For object and event detection, recognition and classification by a robotic vehicle’s perception system, Safran Tech’s teams use supervised learning techniques. “These techniques enable us to develop a system from annotated data. Each sample or data point describes a situation, while the label or annotation associated with it describes how the system is expected to respond in that situation. A system that has learned properly must be able to respond appropriately in situations that it didn’t necessarily encounter during the learning phase.”

In the autonomous systems lab, Yrvann is working on decision mechanisms for autonomous driving systems, using reinforcement learning techniques to refine lateral control of the lab’s robotic vehicles and develop a module enabling them to adjust their speed when entering roundabout intersections.

 

Simulation tools supporting autonomy research

Since arriving at Safran Tech, Yrvann has been maintaining and evolving the digital twin of its robotic vehicles.

“The simulator has a library of sensors for instrumenting simulated vehicles like we do our real platforms, which means we can feed both simulated data and actual sensor data from vehicles into our algorithms,” notes Yrvann. In particular, he tests and validates the decision module building blocks that will define how the autonomous driving system responds on the open road before incorporating them in vehicles.

Special attention also has been given to geolocating road markings, signage and buildings at the Paris-Saclay facility and its environs to within five centimeters in the simulation environment. The simulation tool is vital to the laboratory’s work while also supporting other Safran teams.