Multi-Objective Robotics
Many real-world problems have conflicting objectives, however, it is difficult to design a single reward function that optimally combines all objectives. To address this, we will simultaneously and independently optimize all solutions on the Pareto front. This project will investigate how to implement and improve the existing MO-CMA-ES algorithm to operate a Baxter robot quickly and adaptively in production for a multi-objective problem such as collision avoidance.
- Install Docker and
docker login
- Source
docker.sh
file (Documentation at bottom)
This will run the default NES Algorithm with the parameters in Config.yaml
- Clone this directory
- cd
MOR/
- Run
python train <CONFIG_FILENAME>.yaml
to run the algorithm in the foreground (append an&
at the end to run in the background)- Use
<CONFIG_FILENAME> = "Config"
to run the default Maze example - Check other config files in
cfg/
for other options, or write your own.yaml
config file and add it tocfg/
. - Resolve any dependecy issues that may arise
- Linux/Mac OSX:
sudo -H pip install numpy tensorflow matplotlib pyyaml
- Linux/Mac OSX:
- Use
- Clone this directory
- cd
MOR/
- Run
docker_build mor
- Run
docker_run_link_gazebo mor1 main mor
- Run
python train <CONFIG_FILENAME>.yaml
to run the algorithm in the foreground (append an&
at the end to run in the background)- Use
<CONFIG_FILENAME> = "Config"
to run the default Maze example - Check other config files in
cfg/
for other options, or write your own.yaml
config file and add it tocfg/
.
- Use
- Check the
ext/
directory for your output data - The
.log
file contains the problem state, the reward function, and the results of each individual of each population. - The
.yaml
file contains the parameterws used during training - The
.png
files are graphs of the rewards/success per population.
docker_build <TAG>
: Builds a new container with the given tag namedocker_run <NAME> <TAG>
: Runs the container with the given tag and labels it with the given namedocker_run_link <NAME> main <TAG>
: Same as run above, but links files between host's working directory and the working directory (main
) in the container using a docker volumedocker_run_link_gazebo <NAME> main <TAG>
: Same as above, but with Gazebo enabled between remote and local hosts.docker_inspect <NAME>
: Describes the volume link created by the above commanddocker_stop <NAME>
: Pauses the specified docker containerdocker_exec <NAME>
: SSH's into the containerdocker_rm <NAME>
: Removes a specified containerdocker_ls
: Lists all containersdocker_rm_all
: Removes all containers