Environments

The main object in Gym Anything is GymAnythingEnv.

This is what you use when you want to:

• start an environment
• take actions
• get screenshots and other observations
• finish the run cleanly

The Main Methods

Most code uses these methods:

• reset()
• step(...)
• capture_observation()
• close()

Typical Flow

from gym_anything import from_config

env = from_config(
    "benchmarks/cua_world/environments/moodle_env",
    task_id="enroll_student",
)

obs = env.reset(seed=42)
obs, reward, done, info = env.step([])
env.close()

Fast I/O Observations

Runners with a native fast-I/O backend can opt into in-memory screenshot capture at environment creation time:

env = from_config(
    "benchmarks/cua_world/environments/google_earth_env",
    task_id="take_screenshot",
    fast_io=True,
)

With fast_io=True, screen observations return an in-process image object:

obs = env.capture_observation()
image = obs["screen"]["image"]

The same observation shape is returned by step(...). In fast mode, obs["screen"]["image"] is a PIL image and the library does not write a frame_XXXXX.png file for every observation. QEMU also routes mouse and keyboard actions through QMP and returns immediately after the action plus observation capture, without the legacy fixed post-action sleep. Non-fast mode keeps the existing file-backed obs["screen"]["path"] behavior.

For QEMU, the default fast backend is QMP screendump. To use the lower-latency D-Bus display backend, set:

export GYM_ANYTHING_QEMU_FAST_IO_BACKEND=dbus

See Runners for runner-specific runtime requirements.

Modal Native uses a separate in-VM X11 backend: XDamage and MIT-SHM maintain the latest RGB frame, while XTest injects each keyboard or mouse action as one acknowledged batch. VNC remains available for interactive viewing but is not on the fast-mode screenshot or input path.

What step(...) Does

step(...) sends actions to the environment and returns:

• the new observation
• the reward
• whether the run is done
• extra information in info

If you want to finish a task and run its checker, call:

obs, reward, done, info = env.step([], mark_done=True)

When To Use capture_observation()