Hierarchical Object-Oriented POMDP

Hierarchical Object-Oriented POMDP for Object Rearrangement (2024)

This project addresses the challenge of object rearrangement in partially observable multi-object environments through hierarchical decomposition.

Published at NeurIPS SpaVLE 2025

Key Results

• 71% task completion in partially observable multi-object environments
• 2-3x improvement over baselines through hierarchical decomposition
• Scalability: Handles environments with 20 objects across 4 rooms while baselines limited to 5 objects in 1 room

Technical Approach

• Abstract planner reasons about high-level goals and object dependencies
• Learned low-level policies handle execution and complex constraints
• Object-oriented state representation enables generalization across object types

Challenging Scenarios Handled

• Blocked paths requiring obstacle removal
• Object dependencies (must move A before B)
• Goal conflicts requiring replanning
• Partial observability requiring active exploration

The hierarchical approach enables tractable planning in scenarios where flat approaches fail due to combinatorial explosion.