Train a robot - turn

Train robot to turn a handle

Setup

To train robot with a different behaviour, change values of following coefficients in config file. Values have to be between 0 - 3.

#Coefficients
"coefficient_kw"        : 0,
"coefficient_kd"        : 1,
"coefficient_ka"        : 0,

Reward is splitted into 2 parts, each part is multiplied by specific coefficient and at the end of episode summed up.

Coefficient_kw - not used

Coefficient_kd - is multiplied by distance between robot’s gripper and mid point of predefined line

Coefficient_ka - functional training works without, but You can experiment

Training

  • To train model with default coefficients settings run following command

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

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

Wait until the first evaluation after 250000 steps to check the progress:

training

After 500000 steps the arm is able to switch the lever with 80% accuracy:

training