Train a robot - press

Train robot to press a button

Setup

To train robot with a different behaviour, change values of following coefficients in config file. Values have to be between 0 - 2.

#Coefficients
"coefficient_kw"        : 1.1,
"coefficient_kd"        : 0.65,
"coefficient_ka"        : 0.5,

Reward is splitted into 3 parts, each part is multiplied by specific coefficient and at the end of episode summed up.

Coefficient_kw - multiplies distance between position of robot’s gripper and line - (initial position of robot, final position of robot)

Coefficient_kd - multiplies distance between task_object and robot’s gripper

Coefficient_ka - multiplies position of button

Training

  • To train model with default coefficients settings run following command

python train.py --config ./configs/train_press.json

The training will start with gui window and standstill visualization. New directory is created in the logdir, where training checkpoints, final model and other relevant data are stored.

After 100000 steps check the progress:

training

After 250000 steps the arm is able to switch the lever with 80% accuracy:

training