Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Task Planning for Object Rearrangement in Multi-Room Environments
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a hierarchical task planner for multi-room object rearrangement under partial observability, combining LLM/RoBERTa-based unseen object discovery, CEM-based buffer prediction for swap cases, a directed state graph, and conservative Q-learning, and introduces the MoPOR benchmark.
Problem
Efficiently planning object rearrangement across multiple rooms under partial observability, where unseen objects must be discovered while minimizing agent traversal and steps, and blocked-goal and swap collisions must be resolved without an explicit buffer.
Approach
Interleaves unseen-object search with rearrangement using a commonsense object-room-receptacle predictor (RoBERTa embeddings with filter and ranking networks), resolves swap cases via Cross-Entropy Method buffer search, and plans with a conservative Q-learning agent over a directed spatial graph state.
Key results
- Ours achieves SRN 0.41 and TTL 37.82 for 10 objects with partial occlusion, outperforming heuristic planner (Ours-HP) TTL 45.82 and random search (Ours-RS) SRN 0.19 with TTL 60.72.
- Ours-GE (greedy exploration) fails on fully occluded cases with SRN 0 and EOD not computable.
- Ours maintains consistent SRN and EOD as the number of objects grows from 10 to 30 in MoPOR multi-room scenes.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ground-truth perception instead of the standard perception pipeline (ablation) | Ours · MoPOR multi-room, P.O. and F.O. cases | SRN: 0.41 -> 0.59; 0.39 -> 0.55 | Helped |
| Replace UODM with random receptacle search (ablation) | Ours · MoPOR multi-room, P.O. and F.O. cases | SRN: 0.41 -> 0.19; 0.39 -> 0.13 | Hurt |
| Replace UODM with greedy map-coverage exploration (ablation) | Ours · MoPOR multi-room, P.O. and F.O. cases | TTL: 37.82 -> 70.09; 40.27 -> NC | Hurt |
| Replace the Deep RL planner with a greedy shortest-traversal heuristic planner (ablation) | Ours · MoPOR multi-room, P.O. and F.O. cases | TTL: 37.82 -> 45.82; 40.27 -> 50.27 | Hurt |
Surprising findings
- Ours-GE achieves zero success on fully occluded (inside closed receptacles) cases because its map-coverage exploration does not open receptacles.
Limitations
- Cannot identify unseen objects occluded by clutter on receptacles.
- Assumes perfect motion planning and manipulation capabilities.
- Evaluated only in simulation (AI2-THOR, plus Habitat transfer in video); not validated on real robots.
Experimental setup
Datasets: AMT dataset ProcTHOR apartment scenes iTHOR scenes RoomR
Benchmarks: MoPOR RoomR
Metrics: SRN EOD TTL
Baselines: Ours-GT Ours-RS Ours-GE Ours-HP RoomR
Builds on: Conservative Q-Learning Cross-Entropy Method ProcTHOR AI2-THOR RoBERTa
Closest papers by meaning
- Task Planning for Visual Room Rearrangement under Partial Observability (ICLR 2024 · similarity 0.93)
- Visual Room Rearrangement (CVPR 2021 · similarity 0.89)
- A Category Agnostic Model for Visual Rearrangment (CVPR 2024 · similarity 0.86)
- FloNa: Floor Plan Guided Embodied Visual Navigation (AAAI 2025 · similarity 0.84)
- ManiSkill-HAB: A Benchmark for Low-Level Manipulation in Home Rearrangement Tasks (ICLR 2025 · similarity 0.84)
- REGNav: Room Expert Guided Image-Goal Navigation (AAAI 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.