Scene-Rep-Transformer

This repo is the implementation of the following paper:

Augmenting Reinforcement Learning with Transformer-based Scene Representation Learning for Decision-making of Autonomous Driving
Haochen Liu, Zhiyu Huang, Xiaoyu Mo, Chen Lv
AutoMan Research Lab, Nanyang Technological University
[Paper] [arXiv] [Project Website]

CARLA Environment is now available;
Model Framework Overview:

Get started

1. Download

Clone this repository and navigate to the directory:

https://github.com/georgeliu233/Scene-Rep-Transformer.git && cd Scene-Rep-Transformer

Download required packages:

pip install -r requirements.txt

[NOTE] The requirements lists all the required packages, so you may try to build the simulator first in the case of compatible issues on your devises.

2. Build Scenarios

We keep the independent code strcutures for CARLA and SMARTS, so that you can also choose either of one to install and run:

CARLA

Download all sources of CARLA v0.9.13 via this link
Navigate to envs/carla/carla_env.py , add folder path of the installed CARLA in system PATH in line 18-19:

# append sys PATH for CARLA simulator 
# assume xxx is your path to carla
sys.path.append('xxx/CARLA_0.9.13/PythonAPI/carla/dist/carla-0.9.13-py3.7-linux-x86_64.egg')
sys.path.append('xxx/CARLA_0.9.13/PythonAPI/carla/')

SMARTS

Download & build SMARTS according to its repository
[NOTE] The current scenarios are built upon SMARTS v0.4.18, so you may build from source. Ensure that SMARTS is successfully built.
Download SMARTS Scenarios:

wget https://github.com/georgeliu233/Scene-Rep-Transformer/releases/download/v1.0.0/smarts_scenarios.tar.gz

3. Testing Pipelines

We offered the checkpoints with train_logs for all scenarios:

wget https://github.com/georgeliu233/Scene-Rep-Transformer/releases/download/v1.0.0/data.tar.gz

unzip the ckpts and scenarios:

bash ./tools/download_build.sh

run the scenario test by following example commands:

cd tools
python3 test.py \
        --scenario=left_turn # testing scenarios: [left_turn, cross, carla, ..., etc.]
        --algo=scene_rep # proposed methods

4. Testing Results:

More testing results in [Project Website]

Testing results using different rewards

We adopt two extra reward functions for comprehensive testing:

[R1]; [R2]

Scenario	Left turn			Double Merge			CARLA
Methods	Succ.	Col.	Step(s)	Succ.	Col.	Step(s)	Succ.	Col.	Step(s)
PPO-R1	0.48	0.38	20.4	0.38	0.62	33.7	0.44	0.20	22.1
DrQ-R1	0.70	0.30	27.3	0.66	0.14	38.3	0.74	0.24	17.6
Ours-R1	0.90	0.10	11.7	0.84	0.16	21.2	0.76	0.22	19.4
PPO-R2	0.38	0.62	31.2	0.46	0.54	34.7	0.50	0.20	24.8
DrQ-R2	0.82	0.08	13.9	0.72	0.28	18.7	0.78	0.12	18.5
Ours-R2	0.94	0.04	11.6	0.88	0.10	27.7	0.78	0.16	21.3

Acknowledgements

RL implementations are based on tf2rl

Official release for the strong baselines: DrQ; Decision-Transformer

MarinBao/Scene-Rep-Transformer