🌎 Environment

Description

OmniGibson's Environment class is an OpenAI gym-compatible interface and is the main entry point for interacting with the underlying simulation. A single environment loads a user-specified scene, object(s), robot(s), and task combination, and steps the resulting simulation while deploying user-specified actions to the loaded robot(s) and returning observations and task information from the simulator.

Usage

Creating

Creating a minimal environment requires the definition of a config dictionary. This dictionary should contain details about the scene, objects, robots, and specific characteristics of the environment:

env_simple.py

import omnigibson as og
cfg = {
    "env": {
        "action_frequency": 30,
        "physics_frequency": 60,
    },
    "scene": {
        "type": "Scene",
    },
    "objects": [],
    "robots": [
        {
            "type": "Fetch",
            "obs_modalities": 'all',
            "controller_config": {
                "arm_0": {
                    "name": "NullJointController",
                    "motor_type": "position",
                },
            },
        }
    ]
}

env = og.Environment(configs=cfg)
action = ...
obs, reward, terminated, truncated, info = env.step(action)

Runtime

Once created, the environment can be interfaced roughly in the same way as an OpenAI gym environment, and include common methods such as step, reset, render, as well as properties such as observation_space and action_space. Stepping the environment is done via obs, reward, terminated, truncated, info = env.step(action), and resetting can manually be executed via obs = env.reset(). Robots are tracked explicitly via env.robots, and the underlying scene and all corresponding objects within the scene can be accessed via env.scene.

Types

OmniGibson provides the main Environment class, which should offer most of the essential functionality necessary for running robot experiments and interacting with the underlying simulator.

However, for more niche use-caches (such as demonstration collection, or batched environments), OmniGibson offers the EnvironmentWrapper class to easily extend the core environment functionality.