Interviews, Smart Mobility

Tackling the Dual Problem of Perception and Motion Forecasting – Interview with Aurora’s Principal Software Engineer | Auto.AI USA 2022

Nemanja Djuric, Machine Learning Tech Lead at Aurora

Machine learning approaches to perception are rapidly developing with new models promising higher reliability. To learn more about this, we.CONECT spoke to Nemanja Djuric, Machine Learning Tech Lead at Aurora, who will be speaking at the 5th Auto.AI USA 2022.

we.CONECT: Hi, Nemanja, it is great to have you as our speaker at the 5th Auto.AI USA 2022, and we’d like to ask you a few questions about Aurora’s approach to perception. Let’s start with your role: you are Machine Learning Tech Lead at Aurora; tell us why you are passionate about your job.

Nemanja Djuric: At Aurora, our mission is to deliver the benefits of self-driving technology safely, quickly, and broadly, and all the aspects of that mission come into play when it comes to my role. I am on the Perception team, which develops the capabilities of the Aurora Driver to detect pedestrians, bicyclists, and other objects hundreds of meters away. Perception is also in charge of developing these capabilities for a variety of domains where the Aurora Driver can safely operate in, from long stretches of highway, to airports, to dense urban areas. This broad scope gives me a chance to work on a wide range of topics and problems that we encounter, alongside some of the smartest people in the industry. Working in such a stimulating environment to solve some of the toughest challenges in autonomous driving makes every day an exciting opportunity to learn about something new, which is a privilege and something I truly enjoy.

we.CONECT: At Auto.AI USA, you will be talking about “End-to-end Joint Object Detection and Motion Forecasting Models for Autonomous Driving”. What is the most important message to the participants and your clients in this context?

Nemanja Djuric: Self-driving technology has significantly advanced from when I started in the field seven years ago. Now, we’re using next-generation approaches to development and are seeing results we’d expect from a highly reliable and performant system. The approaches I present show a somewhat different approach and remind us that there are various ways to solve the problems we see on the roads. Ultimately, we always need to keep an eye out for novel ideas that may further improve our systems and help us quickly deliver a safe self-driving product.

we.CONECT: You present the MultiXNet end-to-end model – what makes this model so unique and sets it apart from others being used in the industry?

Nemanja Djuric: In the presentation I discuss the dual problem of perception and motion forecasting of the traffic actors that surround the autonomous vehicle. These are some of the most important problems in the community, which is why combining them under one umbrella using the MultiXNet model is so exciting. This is a departure from the earlier approaches that consider the two problems as separate and try to solve them with two or more separate learned models. We are not the first to propose such an end-to-end, joint solution. However, we propose several extensions and follow-up ideas in my talk that take this technology to the next level in terms of the detection and forecasting performance indicators.

we.CONECT: If you think of your clients’ requirements: Which problem areas regarding autonomous driving do you regard as particularly critical?

Nemanja Djuric: The technology has improved significantly in the last several years, and we see amazing success stories across the industry nearly on a weekly basis. However, the outliers and rare events are something we need to expect and model for. Accounting for situations that are rarely observed and correctly reacting to them remains a critical piece of the puzzle that is not fully there yet.

we.CONECT: What do you think are the most promising technologies to meet these challenges?

Nemanja Djuric: The industry is moving fast to address these challenges and the proposed approaches range from improved learned models that better capture and quantify uncertainty of a traffic scene, simulated environments that can generate rare scenarios before our technology experiences them on the road, improved remote assistance from human operators, and collecting more and more data to help capture the outliers (both on the road and on closed test tracks). A combination of various solutions and approaches is likely to significantly improve the performance of self-driving vehicles in rare scenarios on the road in the near future.

we.CONECT: How do you think will the solutions and technologies for object detection and motion forecasting develop in the next year or two?

Nemanja Djuric: The next few years will be very interesting, as we see significant efforts and progress on many fronts. Software development is moving rapidly, with comprehensive new models and sophisticated approaches published at an amazing rate. We have a plethora of public data sets and leaderboards that are propelling the community to new heights. Hardware improvements are another piece of the puzzle to a self-driving future with improved lidar and camera technologies, more powerful GPUs, and many other advancements progressing at a rapid pace. These will be critical to advancing this industry, and I am really looking forward to seeing what the future brings.

we.CONECT: You are a returning speaker at our events – we are thrilled to have you again this year. What do you expect from Auto.AI USA 2022?

Nemanja Djuric: First of all, I am very grateful to the organizers for inviting me again to share my team’s work! Auto.AI USA is a great place to inform the wider community about the latest progress made and research that we worked on, but also to learn more about the direction that the rest of the industry is headed. Same as earlier times, I am very excited to connect with the other folks from the industry and build connections and relationships for the years to come, that would help us to jointly crack the tough self-driving nut and the challenges that lie ahead of us.

we.CONECT: Thank you for your time, Nemanja, we are looking forward to your session at the conference! 

Nemanja Djuric is presenting a case study End-to-end Joint Object Detection and Motion Forecasting Models for Autonomous Driving on June 20 at 10:45 AM.

Here is what he’ll be talking about:

Object detection and motion forecasting are critical components of self-driving technology, tasked with understanding the current state of the world and estimating how it will evolve in the near future. In the talk we focus on these important problems and discuss end-to-end methods that jointly perform these two tasks, which have shown state-of-the-art performance. We present a deep dive into recently proposed MultiXNet end-to-end model, and discuss a number of improvements that build on this method as a base.

  • learn about the problems of object detection and motion forecasting
  • become familiar with the state-of-the-art approaches that jointly perform these two tasks
  • learn about various modeling improvements that resulted in significant boosts in performance


Auto.AI USA is the leading technical event on deep learning for SAE level 4 and 5 autonomous vehicles bringing together more than 300 top industry experts and decision-makers in machine learning, neural networks, and perception. Join now and discuss self-supervised and behavioral learning concepts, scalable machine and reinforcement learning approaches, benchmarking perception and computer vision systems for ADs with your peers from the automotive AI community.

The 5th Auto.Ai USA, America’s No. 1 event on deep driving for level 4 and 5 autonomous vehicles is back on June 19 – 21, 2022, live in Detroit!

Previous ArticleNext Article