Universal robot control in pybullet environment. Maximum about 200k timesteps were conducted. Because of lack of computational power I couldn't train the agent.
Folder name | Description |
---|---|
best_model |
Best SACstable_baselines3 model |
config |
Config files for training stable_baselines3 and DDPG_torch models |
media |
Media files |
src |
source code |
vid.1.mp4
Episode rewards | Actor loss | Critic loss |
---|---|---|
Reward is based on negative L1 distance between end joint of UR10 robot and target position plus reward for reaching the target position minus constraint for collission with table.
$ pip install git@github.com:Genndoso/UR10_Reach_task.git
Without docker
$ pip install -r requirements.txt
$ python3 main.py --logdir --model_dir --type_of_model [stable, torch] --visualize [True, False] --train [True,False] --config_path --path_to_model
With docker (ex)
$ docker build -t $your_image_name$ --build-arg type_of_model=DDPG_torch --build-arg config_path=config/config_torch.yaml --build-arg train=True --build-arg visualize=True .