The DriveML Challenge

Introduction

Autonomous driving holds the promise of reducing traffic accidents by designing safe, robust, accurate, and intelligent learners. Huawei greatly appreciates this noble goal and is working on researching design and implementation with various European car companies to launch self-driving vehicles as early as 2021 [1]. In particular, Audi has been involved in Huawei Connect 2018, initiating research into Level 4 (L4) automated driving (https://www.huawei.com/en/press-events/news/2018/10/huawei-audi-l4-automatic-driving). In this competition, you are invited to join in the research force of Huawei’s autonomous-driving quest to help develop an intelligent self-driving solution to a variety of emulated traffic scenarios.

At a high-level, autonomous vehicles are typically architected so as to include a perception and a decision-making unit. The perception system covers many sub-tasks such as car localization, static obstacle mapping, moving obstacle detection and tracking, traffic signal recognition, and many others. Based on perception output, the decision making system is responsible for a variety of tasks such as route planning, behavior selection (e.g. change lane, merge, stop at the traffic light), motion control, and execution.

In this competition, you will be given the perception system that our teams have been using, and you are expected to develop your own decision-making system for smart, safe, and efficient driving. The task is a single-agent scenario aimed at controlling one ego vehicle. You will be given six weeks to achieve the tasks.

Why our Competition?

The competition will be conducted in our self-developed simulation engine. Such simulation environment supports both realistic driving behaviors in single and multi-agent cases. We enable flexibility by allowing agents to be trained using hard-coded reinforcement learning policies. Our simulation engine is unique in that it provides three viewing angles of autonomous driving:

Macroscopic

In macroscopic traffic simulation, traffic is modeled at the level of average speed, traffic flow and traffic density. Traffic state is available for a section of the road and not for all individual cars. The method of modelling traffic flow at macroscopic level originated under an assumption that traffic streams as a whole are comparable to fluid streams. The progression of traffic is defined with mathematical equations representing evolution of traffic flow is space and time

macroscopic

Microscopic

In microscopic traffic simulation, each car is modelled individually. At this level of fidelity, the state usually includes the vehicle speed and position along the lane, and lane number when there is more than one lane. These models usually do not have in-lane lateral movement. Two main models control the movement of the car. The car-following model defines the vehicle speed, and lane-change model controls when the car need to change its lane. Also microscopic simulators can better model mixed mode (passenger, public transit, and bicycle, pedestrian) transportation.

microscopic

Mesoscopic

Another level in between is mesoscopic simulation. One of the issues with macroscopic simulation is its inability to track movement of individual vehicles. This is often important when dealing with traffic assignment and traffic routing. When there is a junction, in a macroscopic model, we can only define the percentage of vehicles that go in each direction and cannot differentiate between vehicles commencing from points A or Points B. Mesoscopic simulation attempts to address this issue by tracking individually their state taken from the link they are travelling.

mesoscopic

Our unique simulation

We provide a unique simulation engine that combines the above three levels of granularity. By doing so, our simulation is more realistic, e.g., modelling a whole city where certain areas require higher fidelity. Another reason to have a combinatorial view of simulation is to have realistic traffic flow into the downtown in a dynamic fashion, e.g. in case of accidents on a feeding highway, traffic flow towards the downtown decreases.

We give you the opportunity to work with an accurate, realistic, and scalable simulation engine hoping you help us in figuring out a better driving solution.

Animation

Tasks

Our task will focus on training an RL agent to interact effectively with hard-coded social vehicles. These social vehicles drive safely and continuously along a multi-lane highway, with varied initial conditions for different episodes. Competitors will train a reinforcement learning agent in an environment with the following challenges:

Fig.1 - Lane changing

Fig.2 - Crossing intersections

Fig.3 - Merging

To perform successfully, agents must behave intelligently, with good understanding of how their behaviour influences neighbouring vehicles. For example, agents seeking to overtake a slow car must identify when changing to the adjacent lane is safe, given the presence of fast moving social vehicles.

Observations, Actions and Evaluation

We provide a range of observations to experiment with including:

Fig.1 - Occupancy grid maps, which indicate the positions of nearby vehicles on a grid

Fig.2 - Time-to-collision, which divides the distance between two vehicles in the same lane by their relative velocity

We also provide both low level (throttle, break and steer) and high-level actions (lane-following, target speed, automatic lane change) and well-defined evaluation criteria for both speed and safety. Competitors will need to use the available tools to train policies which optimise the evaluation criteria as assessed on an evaluation task.

Outcomes

Driverless cars must be safe, efficient and robust. Safety involves avoiding collisions and ensuring that the time-to-collision does not get too low, such that other vehicles are required to execute emergency breaking maneuvers (e.g. while merging). Efficiency involves achieving fast and smooth vehicle flow without undue accelerations and decelerations. Robustness means that trained policies should be able to handle a variety of initial conditions. Winning competitors will train policies which excel at all these key features.

Return to homepage