Thanks to the gibson environment developed by StanfordVL, we are able to use that to solve robot motion planning problem.
You can down load the dataset from https://github.com/StanfordVL/GibsonEnv/blob/master/gibson/data/README.md. The folder ' model ' should contain the all the model you download from the link. You should save like ./model/<model name> (e.g. ./model/Allensville)
Please use the following command to install related package:
pip install -r requirement.txt
The file generator.py can display and generate the maps of different floors for all the model. The map will by default saved under the folder of
./[model name]/[floori] (i is the floor number)
You can run the demo and display the map by the following command:
python generator.py
The file map.pywill serve to generate the map for customized number of maps. You can use the following command:
python map.py 1 3 20
To generate the map for 20 models with 1 superposition of binary map and 3 superpostion of raw map.
You will find 3 maps in your <model name>/floor<i> folder and a map_para.txt file containing the map parameters for future use.
The map generated for each floor should look like this:
argv[1]: The number of superposition for binary map,more superposition means better remove the object indoor
argv[2]: The number of superposition for raw map, more superpostion means higher accuracy to describe the model
argv[3](optional): Determining how many models is required to generate maps. The default would run through all the models
File rgb_generator_alpha is used to rendered the rgb image and depth image. For convience, we only consider the case when using the docker
provided by the gibson environment. You can follow the QUICK INSTALLATION(DOCKER) on https://github.com/StanfordVL/GibsonEnv. After going through
the installation on that website, you can start the docker by the following command:
xhost +local:root
docker run --runtime=nvidia -ti --rm -e DISPLAY -v /tmp/.X11-unix:/tmp/.X11-unix -v <host path to model folder>:/root/mount/gibson/gibson/assets/dataset gibson
Then you will get an interactive shell. You have to move rgb_generator_alpha.py to the dir of the model before generating the
image.You can use the following command:
pyhton rgb_generator_alpha.py 50 100 1.2
To generate the image with 50 trajectory for each model and 100 actions in each trajectory. The rgb images and depth images are saved
under the folder named after the starting point of the trajectory in the ./model/<model name>/data folder with a information.txt
file documenting the state information.
argv[1]: The number of trajectory in each map.
argv[2]: The number of steps need to take in each trajectory
argv[3]: the height of the robot
argv[4](optional): Determining how many models is required to go through. In default setting, it would run through all the models.
The rgb and depth image saved under the directory ./model/<model name>/data/<trajectory starting point>/rgb(depth), which should look like:
/rgb/0.png?raw=true "rgb")
To transform the data into tfrecords for training and validation, you have to run the transform_to_TFRecord.py by the following command:
python transform_to_TFRecord.py
The training and validation data will be saved in folder tfdataset.
To display the data stored in ./pfnet/data, run display_data.py by the following command:
python display_data.py ./pfnet/data/<name of tfrecord>
Training requires the training and validation datesets to be downloaded into the ./data/ folder. The configuration of training process is in the configuration file in .pfnet-master/configs/ folder. Note that training requires significant resources, and it may take several hours or days depending on the hardware.You can run the training process by:
cd pfnet-master
python train.py -c ./configs/train.conf --obsmode rgb --mapmode wall