Implementation:Facebookresearch Habitat lab HabitatEnvFactory

Knowledge Sources	Facebookresearch_Habitat_lab
Domains	Embodied_AI, Environment_Management
Last Updated	2026-02-15 00:00 GMT

Overview

HabitatVectorEnvFactory constructs vectorized Habitat environments by splitting scenes across parallel workers, configuring per-environment seeds and measurements, and optionally enabling batch rendering.

Description

HabitatVectorEnvFactory extends VectorEnvFactory and implements the construct_envs method. It loads the dataset to discover available scenes, shuffles them, and distributes scenes across the configured number of environments using a round-robin scheme. When there are fewer scenes than environments, it either reduces the environment count (if enforce_scenes_greater_eq_environments is True) or duplicates scenes across all environments. Each environment receives a unique seed (base seed plus environment index) and has certain rank-specific or rank0-env0-specific measurements filtered out depending on whether the environment is on the first rank. The factory supports both VectorEnv and ThreadedVectorEnv (selected via the HABITAT_ENV_DEBUG environment variable), and initializes a batch renderer if configured.

Usage

Use this factory to create parallel Habitat environments for RL or IL training. Pass a configuration that specifies the number of environments, dataset, and scene settings.

Code Reference

Source Location

• Repository: Facebookresearch_Habitat_lab
• File: habitat-baselines/habitat_baselines/common/habitat_env_factory.py
• Lines: 18-120

Signature

class HabitatVectorEnvFactory(VectorEnvFactory):
    def construct_envs(
        self,
        config: "DictConfig",
        workers_ignore_signals: bool = False,
        enforce_scenes_greater_eq_environments: bool = False,
        is_first_rank: bool = True,
    ) -> VectorEnv:

Import

from habitat_baselines.common.habitat_env_factory import HabitatVectorEnvFactory

I/O Contract

Inputs

Name	Type	Required	Description
config	DictConfig	Yes	Full Hydra/OmegaConf configuration containing habitat and habitat_baselines settings
workers_ignore_signals	bool	No	If True, worker processes ignore OS signals (default: False)
enforce_scenes_greater_eq_environments	bool	No	If True and fewer scenes than environments, reduces the environment count rather than duplicating scenes (default: False)
is_first_rank	bool	No	Whether this is rank 0 in distributed training; controls measurement filtering (default: True)

Outputs

Name	Type	Description
envs	VectorEnv	A vectorized environment with scenes distributed across parallel workers

Usage Examples

Basic Usage

from habitat_baselines.common.habitat_env_factory import HabitatVectorEnvFactory

factory = HabitatVectorEnvFactory()
envs = factory.construct_envs(
    config=my_config,
    workers_ignore_signals=True,
    is_first_rank=True,
)

observations = envs.reset()
# Use envs for training or evaluation
envs.close()