12kobe / carla-aebs

Implementation of an Advanced Emergency braking system in CARLA

Introduction

This repository provides the implementation of an Advanced Emergency Braking System (AEBS) in Carla. We use this AEBS implementation to test the results of the paper "Real time Out of distribution detection in Learning-Enabled Cyber Physical Systems" by Feiyang Cai published in ICCPS'20. This code can be best utilized keeping in mind the following architecture for the AEBS system:

The output from the camera sensor can be used to estimate the distance from the leading car as shown in the architecture. This computed distance can then be input to a learned controller (using the DDPG) algorithm which outputs an appropriate braking signal to control the vehicle velocity.

The scenario that we have implemented in this repo is shown below: We have two actors in the scene (two cars shown in white and red). The car in red is stationary in the scene while the car in white accelerates towards the red car. The objective is that the white car must satisfy the following bounded liveness property:

The white car must stop at a distance of 1-3m from the red car.

Dependency setup

To setup install pipenv using pip install pipenv and run the command pipenv install to setup all the dependencies :)

Running the code:

Collecting data from the simulator

To collect the data from the simulator using the command (in GUI mode):

python collect.py --collect-path /path/where/to/store/images --num-episodes 80 --mode in --gui

To collect the data in a non-gui mode remove the --gui flag

Training the Perception LEC

To train the Perception LEC on the data collected above use the command:

python train_perception.py /path/to/collected/data/ /checkpoint/save/path \
                    --batch-size 64\
                    --random-state 0\
                    --lr 0.01\
                    --optim SGD\
                    --epochs 100

Training the VAE

To train the VAE on the generated data use the command:

python train_vae.py train-vae /path/to/collected/data/ /checkpoint/save/path \
                    --batch-size 64\
                    --random-state 0\
                    --code-size 1024\
                    --lr 0.01\
                    --optim SGD\
                    --epochs 100

Training the RL agent

To train the RL agent using the DDPG algorithm, use the following command (With a GUI. To not use the gui remove the --gui flag):

python aebs.py --save-path /rlagent/chkpt/save/path/ \
                                --num-episodes 1000 \
                                --agent-chkpt-path /chkpt/to/load
                                /paths/from \
                                --gui

Precomputing the calibration scores using the VAE NCM

To compute the calibration scores and save run the method compute_calibration_scores provided in utils/util.py. Running this method would generate a numpy file at the save_path provided to this method!. This can then be used to run the AEBS scenario in an end-to-end fashion

Running the Complete AEBS scenario

To train the RL agent using the DDPG algorithm, use the following command (With a GUI. To not use the gui remove the --gui flag):

python aebs.py --save-path /rlagent/chkpt/save/path/ \
                                --num-episodes 1 \
                                --agent-chkpt-path /path/to/load/rlagent/chkpt/ \
                                --perception-chkpt-path /path/to/trained/perception/chkpt \
                                --vae-chkpt-path /path/to/trained/vae/path/ \
                                --calibration-scores /path/to/precomputed/calibration/scores \
                                --gui --testing --generate-plots

Author

Kushagra Pandey @kpandey008