Wrong shape of actions for vectorized Gymnasium environments with discrete action spaces

[philsteg]

Description

Problem:

Vectorized Gymnasium environments with discrete action spaces fail since SKRL passes actions with a wrong shape to the environments. If the base environment has a discrete action space with shape (), the vectorized version has a multi-discrete action space with shape (n,) (see https://github.com/Farama-Foundation/Gymnasium/blob/main/gymnasium/vector/utils/space_utils.py#L84).

However, SKRL passes actions with the shape (n, 1) to the environment in the step-method of GymnasiumWrapper, and not with the expected shape (n,). This may lead to errors.

Reproduction of bug:

Run the example torch_gymnasium_cartpole_vector_dqn.py

Error:

in gymnasium/envs/classic_control/cartpole.py

lines 165-167: assert self.action_space.contains(action), f"{action!r} ({type(action)}) invalid"

Output:

AssertionError: array([1]) (<class 'numpy.ndarray'>) invalid

What skrl version are you using?

1.4.3

What ML framework/library version are you using?

torch==2.6.0

Additional system information

gymnasium==1.1.1