Auto.AI – Europe’s No. 1 hybrid event on deep learning for Level 4 & 5 autonomous driving.
In the run-up to the event, we.CONECT spoke with Ilan Kadar, Director of AI at Nexar Inc about new technical innovations, latest updates to standards, and pressing challenges regarding Level 4 & 5 autonomous driving.
we.CONECT: What are your main responsibilities in your current role?
Ilan Kadar: Leading the Research AI team and effort to leverage Nexar’s large-scale datasets of real-world driving environments to automotive applications.
we.CONECT: Where do you see the biggest opportunities in the application of AI, machine-, deep- & reinforcement learning in the development of fully autonomous vehicles?
- Scalable solution for collection and indexing of Real-World Corner-Cases for E2E Training (perception, prediction and planning) and Benchmarking of AV’s
- Scalable solutions for real-time and semantic accurate maps for AVs
we.CONECT: How effective are virtual testing, simulation and modelling in Level 4&5 today? What should be improved?
Ilan Kadar: Virtual testing and simulation in Level 4&5 today are very limited due to the lack of real-world corner-cases scenarios (you don’t know what you don’t know)
we.CONECT: What have been the highlights of your work with AI/machine learning in the development of autonomous vehicles so far in your career?
- Safety applications (e.g., ADAS, Driver-Monitoring)
- Scalable tech for building high-quality and fresh maps from unlabeled and noisy data
- Multi-task efficient networks for edge inference (e.g., iOS, Android, Amba)
- 3D reconstruction of collision and corner-cases scenarios from monocular cameras
- Scalable engine for efficient Indexing and retrieval of images and videos (e.g., road-objects, driving scenarios, etc…)
- Continuous learning pipeline for automatic improvement of AI models and adaptation to new environments (e.g., new geo-location, different weather conditions)
- V2V applications (e.g., parking detection, traffic light countdown)
we.CONECT: How is your company developing its AI/machine learning/deep learning capabilities? What are the challenges?
Ilan Kadar: In Nexar we developed a continuous deep-learning pipeline in order to automatically collect corner-cases scenarios and improve/adapt our models to challenging driving scenarios and environments. We are mostly using self-supervised techniques in order to leverage our unique and massive datasets with different modalities (video, IMU, GPS, etc…)
we.CONECT: Please explain in brief the key aspects of your session at the Auto.AI Europe 2021.
Ilan Kadar: In the talk, I will discuss the challenges with AV’s development and deployment at scale and how they can be addressed using crowdsourced vision by leveraging a massive network of smart dash-cams jointly with deep learning models.
we.CONECT: Thank you for the interview! To all our readers, if you are interested and want to join us together with Nexar and want to learn, engage and discuss automotive tech innovation in real-time with thought leaders across the globe, feel free to save your seat: https://www.auto-ai.eu/book-now