Skip to main content

Overview

MansionGym is a Gymnasium-style environment backed by AI2-THOR. You interact with the multi-floor building using a single AI2-THOR-style action API.

Quickstart​

from mansion_gym.mansion_env import MansionGym

env = MansionGym(building_dir="path/to/building", initial_floor=1, max_steps=1000)
obs, info = env.reset()

building_dir should contain per-floor scene JSONs named floor_1.json, floor_2.json, etc.

Step API (AI2-THOR-style only)​

This wrapper accepts only AI2-THOR-style inputs:

# dict form
obs, reward, terminated, truncated, info = env.step({
    "action": "MoveAhead",
    "moveMagnitude": 0.25,
})

# kwargs form
obs, reward, terminated, truncated, info = env.step(
    action="MoveAhead",
    moveMagnitude=0.25,
)

Skills​

Registered skill names: MoveTo, GoNearObject, OpenDoor, CloseDoor, PickupObject, GoToLandmark, CallElevator, TakeStairs, UseElevator, PutObject

See Skills for signatures, parameters, and examples.