The OCAtari Environments

class ocatari.core.OCAtari(env_name, mode='ram', hud=False, obs_mode='ori', render_mode=None, render_oc_overlay=False, *args, **kwargs)

The OCAtari environment. Initialize it to get a Atari environments with objects tracked.

Parameters:

• env_name (str) – The name of the Atari gymnasium environment e.g. “Pong” or “PongNoFrameskip-v5”

• mode (str) – The detection method type: one of raw, ram, or vision, or both (i.e. ram + vision)

• hud (bool) – Whether to include or not objects from the HUD (e.g. scores, lives)

• obs_mode (str) – Define the observation mode. Set to dqn (84x84, grayscaled), ori (210x160x3, RGB image), obj (#Objectsx4). dqn and ori are also organized in a stack of the last 4 frames.

the remaining *args and **kwargs will be passed to the gymnasium.make function.

close(*args, **kwargs)

After the user has finished using the environment, close contains the code necessary to “clean up” the environment. See env.close() for gymnasium details.

property dqn_obs

The 4 (grey+down)scaled last frames (84x84) of the environment, used notably by dqn agents as states.

Type:

torch.tensor

get_ram()

Returns the RAM state

Returns:

The 128 list of RAM bytes

Return type:

list of 128 uint8 values

property get_rgb_state

np.array

Type:

type

property nb_actions

The number of actions available in this environments.

Type:

int

property objects

A list of the object present in the environment. The objects are either ocatari.vision.GameObject or ocatari.ram.GameObject, depending on the extraction method.

Type:

list of GameObjects

property ocstate

A list of the object present in the environment. The objects are either ocatari.vision.GameObject or ocatari.ram.GameObject, depending on the extraction method.

Type:

list of GameObjects

render()

Compute the render frames (as specified by render_mode during the initialization of the environment). If activated, adds an overlay visualizing object properties like position, velocity vector, orientation, name, etc. See env.render() for gymnasium details.

reset(*args, **kwargs)

Resets the buffer and environment to an initial internal state, returning an initial observation and info. See env.reset() for gymnasium details.

set_ram(target_ram_position, new_value)

Directly set a given value at a targeted RAM position.

Parameters:

• target_ram_position (int) – The ram position to be altered

• new_value (int) – The value to put at this RAM position

step(*args, **kwargs)

Run one timestep of the environment’s dynamics using the agent actions. Extracts the objects, using RAM or vision based on the mode variable set at initialization. Fills the buffer if obs_mode was not None at initialization. The observations follow the obs_mode. The method runs the Atari environment env.step() method

Parameters:

action (int) – The action to perform at this step.