Gym: Range of observation and action space in Robots

[Baichenjia]

What is the range of the action space, observation space, desired goal space, achieved goal space?
How can I get it.
For example, in openai/baselines implemention of DDPG, You tell the range of space. But what's the range in Robots enviroment like FetchReach?

class DDPG(object):
def __init__(self, actor, critic, memory, observation_shape, action_shape, param_noise=None, action_noise=None,
gamma=0.99, tau=0.001, normalize_returns=False, enable_popart=False, normalize_observations=True,
batch_size=128, observation_range=(-5., 5.), action_range=(-1., 1.), return_range=(-np.inf, np.inf),
adaptive_param_noise=True, adaptive_param_noise_policy_threshold=.1,
critic_l2_reg=0., actor_lr=1e-4, critic_lr=1e-3, clip_norm=None, reward_scale=1.):

[matthiasplappert]

There are no guarantees for the observation spaces but the actions are always in [-1, 1]. The action space definition should have that set correctly: https://github.com/openai/gym/blob/master/gym/envs/robotics/robot_env.py#L39. You can use env.action_space.low and env.action_space.high to get these ranges. The same works for observations but they will return -np.inf and np.inf, respectively (https://github.com/openai/gym/blob/master/gym/envs/robotics/robot_env.py#L40)

[Jxic]

@matthiasplappert May I ask how I should interpret this action? Is it the change in velocity or joint angles, or is it the desired velocity or joint angles?

[matthiasplappert]

Depends on the task:

• For Fetch, the first 3 dimensions of the action space are an offset in Cartesian space from the current end effector position and the 4th dimension is the state of the parallel gripper in joint space
• For the Shadow hand, all 20 dimensions are the desired joint positions, rescaled such that -1 is the lower limit and 1 is the upper limit

See https://github.com/openai/gym/blob/master/gym/envs/robotics/fetch_env.py#L70 and https://github.com/openai/gym/blob/master/gym/envs/robotics/hand_env.py#L22 for details, respectively.