Principle:Facebookresearch Habitat lab HRL Evaluation
| Knowledge Sources | |
|---|---|
| Domains | Hierarchical_RL, Evaluation |
| Last Updated | 2026-02-15 02:00 GMT |
Overview
Evaluation of hierarchical rearrangement policies measuring composite task success, sub-task completion, and physical interaction quality metrics.
Description
HRL Evaluation extends standard agent evaluation with rearrangement-specific metrics: composite success (all objects placed correctly), individual sub-task success, force metrics (measuring physical realism), and skill utilization statistics. The evaluation loads the full hierarchical policy checkpoint and runs it through rearrangement episodes.
Usage
Use after training a hierarchical policy (or at regular training intervals) to measure rearrangement task performance.
Theoretical Basis
Rearrangement evaluation metrics extend navigation metrics:
- Composite Success: All target predicates satisfied (all objects at goal positions)
- Rearrangement Success: Fraction of objects correctly placed
- Force Metrics: Measure of physical interaction quality (lower collision forces = better)
# Abstract evaluation
composite_success = all(predicate.satisfied(state) for predicate in goal_predicates)
rearrangement_success = sum(p.satisfied(state) for p in goal_predicates) / len(goal_predicates)
Related Pages
Implemented By
Page Connections
Double-click a node to navigate. Hold to expand connections.
Principle
Implementation
Heuristic
Environment