UR10 robot with robotiq_2f_85 gripper

This is a robotic env like gym FetchPickAndPlace-v1 but with UR10 robot arm and robotiq 2f 85 gripper. It uses pybullet simulator to simulate physical interactions between robot and environment.

Generate URDF

To modify URDF model install ros, universal_robot and robotiq
in catkin_ws do:

• source devel/setup.bash
• rosrun xacro xacro ur10_robot_constr.urdf.xacro > my_robot.urdf
• check_urdf my_robot.urdf

Enviroment

Observation:

obs[:3] = gripper_position
obs[3:6] = gripper_orientation
obs[6:9] = gripper_velocity
obs[9:12] = gripper_angular_velocity
obs[12:15] = object_position - gripper_position
obs[15:18] = object_orientation
obs[18:21] = object_velocity
obs[21:24] = object_angular_velocity
obs[24] = gripper_open_value

Action:

action[0:3] = delta x,y,z gripper position
action[3] - open / close gripper

Reward:

dense reward:

reward = 1 - |object_position - desired_position| - |object_position - gripper_position|

space reward:

reward = 0 if goal achieved -1 othervise

Run

evaluate script agent

python3 main.py --agent script --mode eval

visualize script agent

python3 main.py --agent script --mode viz

train ppo agent

python3 main.py --agent ppo --mode train

evaluate ppo agent

python3 main.py --agent ppo --mode eval

visualize ppo agent

 python3 main.py --agent ppo --mode viz

Evaluation results

agent	reward	success
scripted	123.8	0.97
PPO	63	0.1
random	42	0

Video

scripted agent	PPO agent	random agent
scripted	ppo	random