The repo contains the code and benchmark CARLA-ICTS accompanying the master thesis, HyREAL: On Hybrid Learning and Reasoning for Explainable and Safe Navigation of Autonomous Cars in Interactive POMDP. The code is based on the repository: https://github.com/dikshant2210/Carla-CTS02

See here - https://carla.readthedocs.io/en/0.9.13/start_quickstart/ Important: The code is only compatible with CARLA version 0.9.13.

The required python packages can be install with: conda create --name <envname> --file req.yaml. Alternatively, to build an environmemt from scratch, the code is mostly based on tensorflow, pytorch, pandas, matplotlib and scipy.

1. Install a c++ compiler >=7.3.0, g++ is recommended, and make

   sudo apt-get install build-essential sudo apt-get install make

2. Navigate to the following directories (in that order) and run make:

   • ./ISDESPOT/isdespot-ped-pred/is-despot/
   • ./ISDESPOT/isdespot-ped-pred/is-despot/problems/isdespotp_car
   • ./HyLEAP/smart-car-sim-master/is-despot/
   • ./HyLEAP/smart-car-sim-master/is-despot/problems/hybridVisual_car

To train HyREAL-lite from scratch on the whole benchmark execute the following line,

python train_hyreal_a2c.py --server --cuda --port=<port> --config=<config>.

The argument --server specifies to use the carla config for the server, while --cuda specifies that the RL-agent is run on GPU. The carla port is configured using --port and the configuration for the RL-agent is passed via --config.

To test HyREAL-lite for a specific scenario scenario $\in$ [01_int, 02_int, 03_int, 01_non_int, 02_non_int, 03_non_int], run the following command,

python eval_learner.py --server --port=<port> --test=scenario --agent=hyreal

In order to execute all scenarios at once use th following command,

python eval_learner.py --server --port=<port> --test=all --agent=hyreal

To train A2C-CADRL from scratch on the whole benchmark execute the following line,

python train_c2a.py --server --cuda --port=<port> --config=<config>.

The argument --server specifies to use the carla config for the server, while --cuda specifies that the RL-agent is run on GPU. The carla port is configured using --port and the configuration for the RL-agent is passed via --config.

To test A2C-CADRL for a specific scenario scenario $\in$ [01_int, 02_int, 03_int, 01_non_int, 02_non_int, 03_non_int], run the following command,

python eval_learner.py --server --port=<port> --test=scenario --agent=a2c

In order to execute all scenarios at once use th following command,

python eval_learner.py --server --port=<port> --test=all --agent=a2c

To train NavSAC from scratch run,

python3 train_sac.py --server --cuda --port=<port> config=<config>.

To test NavSAC for a specific scenario scenario $\in$ [01_int, 02_int, 03_int, 01_non_int, 02_non_int, 03_non_int], run the following command,

python eval_sac.py --server --port=<port> --test=scenario

In order to execute all scenarios at once use th following command,

python eval_sac.py --server --port=<port> --test=all

HyLEAP can be trained by running,

python train_hyleap.py --port=<port>.

The parameters can be adjusted in hyleap/hybrid_visual.py

To test HyLEAP for a specific scenario scenario $\in$ [01_int, 02_int, 03_int, 01_non_int, 02_non_int, 03_non_int], run the following command,

python eval_hyleap.py --port=<port> --despot_port=<despot_port> --test=scenario

The argument --desport_port specificies the port over which the RL-agent communicates with the planner. To run the evaluation for all scenarios the following command can be used,

python eval_hyleap.py --port=<port> --despot_port=<despot_port> --test=all

For IS-DESPOT(*) no training is required. The evaluation for scenario $\in$ [01_int, 02_int, 03_int, 01_non_int, 02_non_int, 03_non_int] can be run via

python eval_isdespot.py --server --port=<port> --despot_port=<despot_port> --test=scenario python eval_isdespot_star.py --server --port=<port> --despot_port=<despot_port> --test=scenario

To run, the evaluation for all scenarios run:

python eval_isdespot.py --server --port=<port> --despot_port=<despot_port> --test=all python eval_isdespot_star.py --server --port=<port> --despot_port=<despot_port> --test=all

In order to train P3VI it is required to first extract pedestrian paths from the simulation. For this the script exctract_pp_data.py can be used. It is advised to extract the interactive and non-interactive scenarios seperately to reduce the runtime. E.g. the command,

python extract_pp_data.py --server --port=<port> ,

extracts the non-interactive scenarios and the command,

python extract_pp_data.py --server --port=<port> --int ,

extracts the interactive scenarios. Next, got to the directory P3VI and adjust the paths in the file train.py. The path predictor can be trained with the command,

python train.py.

Shield:

This work is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.