prcdmp
DMP formulation after "Learning and generalization of motor skills by learning from demonstration", Pastor et. al 2009 (https://ieeexplore.ieee.org/abstract/document/5152385)
setup
The python part of this project uses the environment pydmp (specified in pydmp.yml). Using Anaconda (https://www.anaconda.com/distribution/#download-section) you can easily set up this environment using the following command in the source directory of this repo:
conda env create -f pydmp.yml
configuration
Trajectories are organized in folders for each trajectory in data/. In data/nameOfTrajectory/conf/dataset.json you can configure the following:
| Field | Description |
|---|---|
| robot_ip | IP of the robot, needs to be reachable by the machine running the code |
| current_episode | recording a new trajectory in the same folder will increment this counter |
DMP specific configuration is handled in data/nameOfTrajectory/conf/DMP.json. Important fields are the following:
| Field | Description |
|---|---|
| q0 | joint space initial configuration of the robot |
| goal | joint space goal configuration of the robot (will be adapted accordingly after recording |
| n_basis | amount of basis functions to approximate the trajectory |
| timespan | duration of trajectory for recording in [s] |
| dofs | degrees of freedom |
| dt | timestep size in [s] |
usage
recording trajectories
The script record.sh is for recording a new trajectory with the robot specified in conf.json.
Call ./record.sh <nameOfTrajectory> w to record a trajectory with the specified name. The w issues the weight fitting, without it, only the trajectory will be saved.
replaying trajectories
The relevant script is rundmp.sh. Again the name of the trajectory you want to run needs to be specified. Also the current_episode needs to be set accordingly (always one higher than what you want to see, so set it to 1 if you want to run 0, yes, weird, but a leftover I did not bother fixing)
Type ./rundmp.sh <nameOfTrajectory> r for the trajectory to be executed on the robot.
Acknowledgement
The python part of the project is based on work from this repository: https://github.com/studywolf/pydmps Thanks also for the great resources in the corresponding blog: https://studywolf.wordpress.com/2013/11/16/dynamic-movement-primitives-part-1-the-basics/