NVIDIA GeForce RTX 3060, python3.6, torch1.10.2+cu113
-
First install Anaconda
-
Use conda environment
conda env create --file robustnav.yml --name <ENV-NAME> e.g. conda env create --file robustnav.yml --name rdmaenav -
Activate the conda environment
conda activate <ENV-NAME>
Specify the path of the pretrained visual module in "allenact/embodiedai/preprocessors/vit.py "
e.g.
pretrain_path = "/home/pengjie/projects/RDMAE_Nav/checkpoints/rdmae.pth"
Run the following command to train a navigation agent
bash train.sh
python main.py \
-o storage/robothor-pointnav-rgb-rdmae-ddppo \
-exp pointnav_robothor_vanilla_rgb_rdmae_ddppo \
-b projects/robustnav/experiments/train \
-s 12345 \
-et rnav_pointnav_vanilla_rgb_rdmae_ddppo_clean
#Parameters Description
-o: Model and log storage location
-exp: Experimental configuration, together with -b to form the task file and path
-b: Task folder path
-s: Random seed
-et: The name of the folder where the tensorboard is saved
The pretraining model consists of two parts：visual module and policy module
-
Specify the path of the pretrained visual module in "allenact/embodiedai/preprocessors/vit.py "
e.g. pretrain_path = "/home/pengjie/projects/RDMAE_Nav/checkpoints/rdmae.pth" -
Specify the path of the pretrained policy module in "test.sh"
e.g. agent=/home/pengjie/projects/RDMAE_Nav/checkpoints/rdmae_nav.pt
Run the following command to train a navigation agent
bash test.sh
Tests for clean and visual corruptions. Results are saved in "storage/robothor-pointnav-rgb-rdmae-ddppo-eval".
Specify the episodes to be visualized in projects/tutorials/running_inference_tutorial.py, running_inference_tutorial_fov.py, and running_inference_tutorial_cam_crack.py.
viz_ep_ids: Specify the episodes to be visualized in the top-down trajectory
viz_video_ids： Specify the episodes to be visualized for egocentric navigation
bash viz.sh
The visualization results are saved in the tensorboard file in "storage/robothor-pointnav-rgb-rdmae-ddppo-eval"
egocentric navigation
top-down trajectory
This project requires internet connection to render the UI interface when running, so internet connection is required during training/testing and can be disconnected after training/testing starts.
This project has heavily referenced code from robustnav, thanks!
This project was completed by Daniel Peng.