This repository contains the codes for our International Journal of Social Robotics paper: Crowd-comfort Robot Navigation among Dynamic Environment Based on social-stressed deep reinforcement learning (ssDRL)

Our paper is available online Now! Crowd-Comfort Robot Navigation Among Dynamic Environment Based on Social-Stressed Deep Reinforcement Learning

Robot navigation in a dynamic environment needs to consider the comfort of the surrounding pedestrians under the premise of ensuring the safety of human, which is a challenging task. This paper proposes the concept of social stress index based on tension space of robot and human, which is an important part of Human-Robot interaction. Especially, the proposed approach develops crowd-comfort navigation by combining social stress index with a deep reinforcement learning framework and the value network. A set of typical simulation experiments show that our method effectively improves the comfort of surrounding pedestrians during the process of robot navigation.

• Python-RVO2 library
• Python>=3.5.0
• Pytorch>=1.5.0
• torchvision>=0.6.0
• opencv3>=3.1.0
• numpy>=1.14.2

We set up two different environments, the first is a simple mode (env.config) and the second is a difficult mode.

• Simple mode: only 5 persons in scene and the sensing range of the robot is not considered
• Difficult mode : 10/20 persons in scene and the sensing range of the robot is set to 3m

We first train in simple mode and then fine-tuned on the difficult mode

python train.py --policy ssdrl --train_stage 1 --gpu True

After get model parameters file rl_model.pth in data/output then

python train.py --policy ssdrl --train_stage 2 --gpu True

For evaluation in stage one (simple mode) use --visualize True can visualize the case

python test.py  --train_stage 1 --gpu True --visulaize True

For evaluation in stage one (simple mode)

python test.py  --train_stage 2 --gpu True

In this project, some codes for environment simulation and evaluation are built upon ICRA2019-Crowd-Robot Interaction: Crowd-aware Robot Navigation with Attention-based Deep Reinforcement Learning