📃 Confidence-Guided Human-AI Collaboration: Reinforcement Learning with Distributional Proxy Value Propagation for Autonomous Driving

🔥 Source Code Released! 🔥

[arXiv]

1. This work introduces Distributional Proxy Value Propagation (D-PVP), which integrates human intention into distributional reinforcement learning, enabling efficient policy learning with minimal human intervention.

2. A shared control mechanism and policy confidence evaluation algorithm dynamically balance human-guided and self-learning policies, ensuring both safety and performance in autonomous driving.

3. The proposed method is validated in both MetaDrive and real-world urban driving using a sensor-equipped UGV. Extensive experiments demonstrate superior performance in terms of sample efficiency, safety, and generalization across diverse traffic scenarios.

Email: lizeqiao@tju.edu.cn

Framework

Demonstration

Training example using C-HAC

Testing example

C-HAC Real-World Driving Demonstration – Route 1

C-HAC Real-World Driving Demonstration – Route 2

User Guide

Clone the repository.

cd to your workspace and clone the repo.

git clone https://github.com/lzqw/C-HAC.git

Create a new Conda environment.

cd to your workspace:

conda create -n CHAC python=3.9

Activate virtual environment.

conda activate CHAC

Install Pytorch

Select the correct version based on your cuda version and device (cpu/gpu):

pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118

Install other requirements.

# Install the requirements.
pip install -r requirements.txt

Training

Modify the sys path in example_train file, and run:

python train_dsact_pvp_rl.py