Social navigation with SAC_AE algorithm

EmoBipedNav: Emotion-aware Social Navigation for Bipedal Robots with Deep Reinforcement Learning

• video, project-website
• the networks used in this repo are partially from sac-ae

environment

• Ubuntu 22.04
• NVIDIA GeForce RTX 4090
• Driver Version: 535.171.04 CUDA Version: 12.2
• Anaconda install
• Create Anaconda env, conda create -n torch python=3.10. Do NOT use python 3.12 because pybind11 is not compatible when you compile pybind project.
• conda activate torch
• Install pytorch

simulation with simplified LIP model and digit_mujoco

• git clone https://github.com/sybrenstuvel/Python-RVO2.git

• conda activate torch

• pip install Cython torchviz tensorboard

• inside Python-RVO2, python setup.py build, and then python setup.py install.

• inside C_library, python setup.py build_ext --inplace

• in digit_mujoco, pip install -e ., installing low-level controller for Digit

• train and evaluate the model with digit_mujoco, python emobipednav_main.py .

• test mode, python emobipednav_main.py --load_test_model YOUR_MODEL_PATH, YOUR_MODEL_PATH should be follow the format like step_100000_success_90.

• By default, training, evaluation, and test are all based on Digit in MuJoCo. You can use LIP by changing the parameter robot_model, robot_eval_model, and robot_test_model.

• replay saved trajectory, python replay_episode.py, please revise file path in ./logs/XXX

benchmark training and testing

• dwa, python dwa_main.py
• drl-vo, python drl_vo_main.py
• lidar-sac, python lidar_sac_main.py
• lndnl, python lndnl_main.py
• using only a single lidar grid map, python emobipednav_main.py --frame_stack 1
• using occupation grid map, python emobipednav_main.py --grid_map
• for the tests of drl-vo, lidar-sac, and lndnl, python xxx_main.py --load_test_model YOUR_MODEL_PATH

replay episode

• python replay_episode.py, please the the file path inside this script