Skip to content

Honda R&D publishes first results of Autonomous Emergency Steering trained using DeepSim

The “Innovative Research Excellence” division of Honda R&D recently published their work on the development of an autonomous emergency steering system, developed with the DeepSim platform. Honda R&D wrote up their experiences in the paper called “Autonomous Emergency Steering Using Deep Reinforcement Learning for Advanced Driver Assistance System” and presented this work at the 59th Annual Conference of the Society of Instrument and Control Engineers of Japan (SICE). We are proud to summarize this cutting-edge work for you in this post.

This system (see Fig 1.) is built around a path planning module that uses Reinforcement Learning. The ADAS sensors feed the perception modules which send their observations to the path planning module which uses a neural network trained with DeepSim. The neural network analyzes the incoming data and determines the best action to take. This action (car’s path) is then sent to the vehicle control module which operates the car. This entire system runs embedded within the car and updates the vehicle operations multiple times per second.

Fig 1. Autonomous Emergency Steering System Diagram

The path planning module is trained completely in a virtual world. This is done via a custom driving simulator, developed jointly by and Honda R&D. The simulator is connected to DeepSim and uses reinforcement learning to train an automatic collision avoidance system. Multiple random scenarios are generated in order to create a robust system that can determine the best course of action in varying conditions with respect to speed, distance, surface area, etc.

The generated scenarios serve as input for the training process and DeepSim performs the cloud based distributed reinforcement learning training. Once the training process is complete the optimized path planning module is exported and optimized for the embedded deployment platform using the DeepSim deployment tools.

Honda ran deployment tests on a closed track using their test vehicle and recreating real versions of the virtual scenarios (see Fig. 2). The tests showed that the agent was able to avoid a collision with both a scooter that would appear from the left side of the road as well as a pedestrian walking in the opposite lane. With help of the DeepSim platform Honda has succeeded in creating a controller robust enough to execute demo scenarios in the real world.

Honda will continue this work by increasing the number of training scenarios in order to develop more advanced agents that can automatically prevent collisions when deployed in the next generation of road cars.

The paper can be obtained from:

For more information about this project or the DeepSim platform please contact us:

Fig 2: Two examples scenarios. The scooter is hidden behind the wall and will appear at random. In the left panel the Ego car can safely move to the right lane in order to avoid the scooter. In the right panel the Ego car must find a more complicated path since the pedestrian must also be avoided.


How can we help?

Reach out – we’d love to hear about how we can help.

We use cookies and similar technologies to enable services and functionality on our site and to understand your interaction with our service. By clicking on accept, you agree to our use of such technologies for analytics. See Privacy Policy

Leave this field blank