Principle:Facebookresearch Habitat lab HRL Configuration Composition
| Knowledge Sources | |
|---|---|
| Domains | Configuration_Management, Hierarchical_RL |
| Last Updated | 2026-02-15 02:00 GMT |
Overview
Configuration composition specialized for hierarchical RL experiments, defining skill policies, high-level policy selection, and skill checkpoint paths.
Description
HRL Configuration Composition extends the standard Hydra config composition with hierarchical-policy-specific sections. The config must define: which skills are available (defined_skills mapping), what type of high-level policy to use (neural, fixed, or planner), paths to pre-trained skill checkpoints, and per-skill observation/action space mappings.
This configuration layer enables composing different skill combinations and high-level policy strategies without code changes.
Usage
Use when setting up hierarchical RL training for rearrangement tasks. The HRL config inherits from the base PPO config and adds hierarchical policy definitions.
Theoretical Basis
The hierarchical config defines a two-level policy structure:
- Skill definitions: Each skill maps a name to a policy class, observation mask, action sub-space, and optional checkpoint path
- High-level policy: Selects which skill to execute at each decision point
- Termination conditions: Define when each skill hands control back to the high-level policy
Config structure (abstract):
# Hierarchical config structure
hierarchical_policy:
high_level_policy:
name: "hl_neural" # or "hl_fixed", "hl_planner"
defined_skills:
nav_to_obj:
skill_name: "NavSkillPolicy"
checkpoint_path: "checkpoints/nav.pth"
pick:
skill_name: "PickSkillPolicy"
checkpoint_path: "checkpoints/pick.pth"