Universal robot control in pybullet environment with reinforcement learning

Universal robot control in pybullet environment. Maximum about 200k timesteps were conducted. Because of lack of computational power I couldn't train the agent.

Project structure

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

Results

vid.1.mp4

Episode rewards	Actor loss	Critic loss

Reward design

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.

$$ -\sum_{k=1}^3 joint_k - tp_k + gr - collission $$

Install

$ pip install git@github.com:Genndoso/UR10_Reach_task.git

Usage

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 .