Implementation:Facebookresearch Habitat lab HierarchicalPolicy init

Knowledge Sources	Habitat-Lab
Domains	Hierarchical_RL, Software_Architecture
Last Updated	2026-02-15 02:00 GMT

Overview

Concrete constructor for the hierarchical policy that loads and orchestrates pre-trained skill sub-policies under a high-level controller, provided by habitat-baselines.

Description

The HierarchicalPolicy.__init__ method reads the defined_skills config mapping, instantiates each skill policy (loading checkpoints via NnSkillPolicy.from_config), creates the high-level policy (neural, fixed, or planner), and sets up the action/observation space routing. The act method implements the option-framework execution: the high-level policy selects a skill, the skill executes until termination, then control returns.

Usage

Instantiated by the registry during PPOTrainer._init_train() when the policy config specifies HierarchicalPolicy.

Code Reference

Source Location

• Repository: habitat-lab
• File: habitat-baselines/habitat_baselines/rl/hrl/hierarchical_policy.py
• Lines: L41-122 (__init__), L280-392 (act)

Signature

class HierarchicalPolicy(nn.Module, Policy):
    def __init__(
        self,
        config,
        full_config,
        observation_space: spaces.Space,
        action_space: ActionSpace,
        orig_action_space: ActionSpace,
        num_envs: int,
        aux_loss_config,
        agent_name: Optional[str] = None,
    ):
        """
        Args:
            config: Hierarchical policy config section
            full_config: Complete experiment config
            observation_space: Full environment observation space
            action_space: Mapped action space for this policy
            orig_action_space: Original environment action space
            num_envs: Number of parallel environments
            aux_loss_config: Auxiliary loss configuration
            agent_name: Optional agent name (for multi-agent)
        """

Import

from habitat_baselines.rl.hrl.hierarchical_policy import HierarchicalPolicy

I/O Contract

Inputs

Name	Type	Required	Description
config	DictConfig	Yes	Hierarchical policy config with defined_skills and high_level_policy
full_config	DictConfig	Yes	Complete experiment config
observation_space	spaces.Space	Yes	Environment observation space
action_space	ActionSpace	Yes	Mapped action space
num_envs	int	Yes	Number of parallel environments

Outputs

Name	Type	Description
policy	HierarchicalPolicy	Assembled hierarchical policy with loaded skills and HL policy

Usage Examples

Construct Hierarchical Policy

from habitat_baselines.rl.hrl.hierarchical_policy import HierarchicalPolicy

# Typically constructed by registry during training init
policy = HierarchicalPolicy(
    config=config.habitat_baselines.rl.policy.hierarchical_policy,
    full_config=config,
    observation_space=envs.observation_spaces[0],
    action_space=envs.action_spaces[0],
    orig_action_space=envs.orig_action_spaces[0],
    num_envs=config.habitat_baselines.num_environments,
    aux_loss_config=config.habitat_baselines.rl.auxiliary_losses,
)

Related Pages

Implements Principle

• Principle:Facebookresearch_Habitat_lab_Hierarchical_Policy_Assembly

Requires Environment

• Environment:Facebookresearch_Habitat_lab_CUDA_GPU_Training_Environment