Brain-Inspired Deep Imitation Learning for Autonomous Driving Systems

Project Abstract

Autonomous driving has attracted great attention from both academics and industries. To realise autonomous driving, Deep Imitation Learning (DIL) is treated as one of the most promising solutions, because it improves autonomous driving systems by automatically learning a complex mapping from human driving data, compared to manually designing the driving policy. However, existing DIL methods cannot generalise well across domains, that is, a network trained on the data of source domain gives rise to poor generalisation on the data of target domain. We propose a novel brain-inspired deep imitation method that builds on the evidence from human brain functions, to improve the generalisation ability of DNN so that autonomous driving systems can perform well in various scenarios. Specifically, humans have a strong generalisation ability which is beneficial from the structural and functional asymmetry of the two sides of the brain. We design dual Neural Circuit Policy (NCP) architectures in DNN based on the asymmetry of human neural networks. Experimental results demonstrate that our brain-inspired method outperforms existing methods regarding generalisation when dealing with unseen data.

Note: The instructions described below are related to Windows OS. They are pretty straightforward given that the user has already downloaded the necessary project files and meets the dependencies and requirements mentioned in maintenance_manual.pdf.

Data Preprocessing

The comma.ai dataset is split into 3 types of raw recordings prior to being preprocessed. This process is carried out manually and no script is developed for the purpose. We present the actions required to preprocess these 3 types of recording data. In contrast, the Udacity dataset is preprocessed on the go while training the models via a batcher function thus no additional steps are required.

• Step 1. Enter a command prompt terminal window;

• Step 2: Navigate to the directory of the comma.ai dataset model training scripts by using the cd command;

  • Alternatively, one could go to the model training program files directory via Windows File Explorer, use the Ctrl+L keyboard shortcut to go to the address bar and enter cmd, as shown in Figure A.1, to launch a command prompt from the current file location;

• Step 3: Run the data preprocessing script by entering the following command line arguments into the terminal:

  python preprocess_data.py rec_type [batch_size]

  • The rec_type should be set to the value of either sunny, cloudy or night since we have 3 types of video files. On the other hand, the optional batch_size argument above can be replaced with the desired numerical size value of each batch. It is set to 64 by default.

You can skip the above steps and directly download the 64-batched, preprocessed dataset here.

For additional details about the project (how to reproduce, dependencies etc.), please refer to the maintenance_manual.pdf and user_manual.pdf files.