README.md

Panda-Gym

• The Panda-Gym repo: https://github.com/qgallouedec/panda-gym/
• The MiniGrid document: https://panda-gym.readthedocs.io/en/latest/index.html

Environment Overview

There are two control modes:

• end-effector control: "ee" envs: the action is the displacement of the end-effector.
• joint control: "joint" envs: the action is the motion of the joints.

Observations

By default, the observation has three fields:

• "observation": 0-5: robot observation, 6-8: object position, 9-11: object velocity
• "achieved_goal": 0-2: object position
• "desired_goal": 0-2: goal position

Actions

• For the "ee" envs: Box(-1,1,(3,),float32).
• For the "joint" envs: Box(-1,1,(7,),float32).

Rewards

• Sparse: the environment return a reward if and only if the task is completed.
• Dense: the closer the agent is to complete the task, the higher the reward.

Important

We need to modify the core.py script in the panda-gym library, at line 236:

desired_goal_shape = observation["desired_goal"].shape

and the line 240-241:

desired_goal=spaces.Box(-10.0, 10.0, shape=desired_goal_shape, dtype=np.float32),
achieved_goal=spaces.Box(-10.0, 10.0, shape=achieved_goal_shape, dtype=np.float32),