Speaker Interviews
Utilizing deep learning for AV simulation
Cognata
Danny Atsmon, CEO of Cognata, outlines the long (virtual) road ahead for simulation to truly guarantee safe AVs that can successfully deal with any situation, or at least be at the same level as a human driver.
Catch Danny’s presentation ‘Pushing the boundaries – simulation vs. the real world’ at the Autonomous Vehicle Test & Development Symposium. Purchase your delegate pass here.
Tell us more about your presentation.
One of the biggest challenges of the AV industry is training and validating the autonomous driving cores through a large number of tests. According to the think tank RAND corporation, it will take AV makers roughly 11bn miles to bring an autonomous software stack to a level at which it could act and drive as safely as a human being. Simulation software will play a major role in accelerating the development of ADAS and AV. Beyond the issue of scale, the fidelity of the data and its resemblance to the real world (level of realism) is a key factor in reducing the need for expensive physical road tests and dramatically cutting the time-to-market.
In my talk, I will present a mathematical matrix that offers a quantitative way of assessing the simulation performance compared to real data and I will show how a proper simulation can be constructed based on deep learning techniques.
What are the advances you've made in the field of AI?
Cognata is transforming real-life data into simulation assets by building the ‘digital twin’ of the world, including real-life traffic models which are based on analysis of camera streams. In addition, we use sensor emulation that is based on actual sensor output. Cognata uses AI, unsupervised learning and deep neural networks to complement these advanced transformations and technological innovations.
What challenges are still to be solved in simulation?
The main challenge that we are currently facing is to capture as many edge cases as possible and to do this at increasing speed, while also scaling the operation to almost-infinity. That means running each scenario across all possible permutations, with various traffic conditions, different weather conditions and different driving cultures.
To achieve this, we need to be able to create a very realistic environment, replicate different driving cultures based on geographical traffic patterns, and present very high sensor and camera emulation capabilities. In addition, this needs to be done at massive scale while remaining purely artificial. It’s a colossal task.