This repo is accompanying our paper "Delay Aware Universal Notice Network : real world multi-robot transfer".
- Python 3.7
- ml-agents 0.15.1
- ml-agents-envs 0.15.0
- TensorFlow 1.15.2 or higer
Install the python ml-agents package (this will also install TensorFlow and ml-agents-envs) and gdown with
python -m pip install mlagents==0.15.1 pip install gdown in your favorite virtual environment.
Clone this repository or dowload the zip file and unzip it.
All the bash commands below need to be lauch from the root of the Unity directory
cd Unity/To download the required Unity environments, run
python downloadEnv.pyat the root of the cloned github repo. It will create a directory called env. (If you are on Linux/OSX, you may need to add the execution right for the environments).
mkdir log_outputto create the directory where the log_output will be stored (on OSX create this directory inside the env directory). It will not run without this directory.
For windows user, you must add /gutter_task to all env argument for the below commands, e.g
--env=env/BAM_training/gutter_taskTo train the delay aware UNN module on the BAM robot run
mlagents-learn config/config_gutter.yaml --run-id=BAMDelayAware --train --env=env/BAM_training --width=756 --height=756 --time-scale 10 --env-args --aware 0To train the delay unaware UNN module on the BAM robot run
mlagents-learn config/config_gutter.yaml --run-id=BAMDelayUnaware --train --env=env/BAM_training --width=756 --height=756 --time-scale 10 --env-args --aware 1To test the delay aware UNN module after training on the other simulated robots run
mlagents-learn config/config_gutter.yaml --run-id=BAMDelayAware --load --env=env/ARG_ROBOT --width=756 --height=756 --time-scale 1 --env-args --delay ARG_DELAY --aware 0 To test the delay unaware UNN module after training on the other simulated robots run
mlagents-learn config/config_gutter.yaml --run-id=BAMDelayUnaware --load --env=env/ARG_ROBOT --width=756 --height=756 --time-scale 1 --env-args --delay ARG_DELAY --aware 1with ARG_ROBOT = BraccioUNN_transfer for the braccio robot or ARG_ROBOT = 2DoFUNN_transfer for the 2 DoF robot
where ARG_DELAY in [0.1 , 1] is the delay of the system
To train the delay aware Vanilla agent on the robot run
mlagents-learn config/config_gutter.yaml --run-id=(ARG_ROBOT)DelayAware --train --env=env/ARG_ROBOT --width=756 --height=756 --time-scale 3 --env-args --aware 0To train the delay unaware Vanilla agent on the robot run
mlagents-learn config/config_gutter.yaml --run-id=(ARG_ROBOT)DelayUnaware --train --env=env/ARG_ROBOT --width=756 --height=756 --time-scale 3 --env-args --aware 1with ARG_ROBOT = VanillaBraccio for the braccio robot or ARG_ROBOT = Vanilla2DoF for the 2 DoF robot
UNN module on the BAM robot
UNN module on the Braccio robot
UNN module on the 2DoF robot
- Python 2.7
- ROS Kinetic
- TensorFlow 1.15.2 or higer
- OpenCv
Once the models are fully trained on the simulated environments, the tensorflow model will be located inside model/run-id/My Behavior/frozen_graph_def.pb
To launch the ROS master run
roscoreAll the bash commands below need to be lauch from the root of the ROS directory
To try the models on the real robots, first launch the camera.py script which will track the required poses a
python camera.pyPls note that the script expect the ball to be green and a yellow circle at one of the end of the gutter to get the ball position and velocity.
To run the UNN based models :
python unn.py -p ARG_PATH -m ARG_METHOD -r ARG_ROBOT -t ARG_TESTor the Vanilla models run :
python vanilla.py -p ARG_PATH -m ARG_METHOD -r ARG_ROBOT -t ARG_TESTwhere ARG_PATH is the path to the tensorflow model ARG_METHOD is an integer indicating wether the model is delay aware (0) or not (1) ARG_ROBOT is an integer specifying the type of robot to use : braccio(0) or 2DoF(1) ARG_TEST is an integer specifying the type of test to run : short test (0) or long test (1). A short test is composed of 3 desired ball position whereas a long test is composed of 50.
UNN module on the Braccio robot
UNN module on the 2DoF robot