Research map / Point clouds / Point cloud classification
TP-MDDN: Task-Preferenced Multi-Demand-Driven Navigation with Autonomous Decision-Making
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.
Introduces TP-MDDN, a long-horizon multi-sub-demand navigation benchmark with task preferences, and AWMSystem (BreakLLM, LocateLLM, StatusMLLM) with MASMap, Dual-Tempo action generation, and an Adaptive Error Corrector for training-free navigation in AI2-THOR/ProcTHOR.
Problem
Demand-driven navigation handles only one need at a time; real tasks involve multiple sequential needs with personal task preferences, requiring long-horizon navigation with subtask tracking.
Approach
A modular, training-free system where LLMs decompose instructions, select goals from object memory, and monitor subtask completion, combined with a dual-tempo controller that alternates slow LLM-guided A* planning and a fast pretrained policy, plus a 3D-to-2D accumulated semantic map and error correction.
Key results
- AWM-Nav achieves 32.00% SR and 62.89% ISR on TP-MDDN, versus 16.00% SR for DDN and InstructNav and 13.33% for MO-DDN.
- Average execution time per long-horizon instruction is 6.82 minutes for AWM-Nav versus 88.90 minutes for InstructNav.
- Swapping RAM-Grounded-SAM for GLEE or YOLO lowers SR from 32.00 to 21.33 and 29.33 respectively.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use AWM-Nav instead of DDN | TP-MDDN navigation · TP-MDDN | SR: 16.00 -> 32.00 | Helped |
| use AWM-Nav instead of InstructNav | TP-MDDN navigation · TP-MDDN | SR: 16.00 -> 32.00 | Helped |
| replace RAM-Grounded-SAM with GLEE (ablation) | AWM-Nav · TP-MDDN | SR: 32.00 -> 21.33 | Hurt |
| replace RAM-Grounded-SAM with YOLO (ablation) | AWM-Nav · TP-MDDN | SR: 32.00 -> 29.33 | Hurt |
| use Qwen2-5-VL-72B instead of Qwen2.5-VL-7B (ablation) | AWM-Nav reasoning · TP-MDDN | SR: 19.33 -> 32.00 | Helped |
| use Qwen2-5-VL-72B instead of GPT-4o (ablation) | AWM-Nav reasoning · TP-MDDN | SR: 28.67 -> 32.00 | Helped |
| remove BlockCorret (ablation) | AWM-Nav Adaptive Error Corrector · TP-MDDN | SR: 32.00 -> 27.33 | Hurt |
| remove BeyondCorret (ablation) | AWM-Nav Adaptive Error Corrector · TP-MDDN | SR: 32.00 -> 28.00 | Hurt |
| add StatusMLLM for task-status tracking (ablation) | AWM-Nav · TP-MDDN | SR: 27.33 -> 32.00 | Helped |
Surprising findings
- GPT-4o as the reasoning model gives lower SR (28.67) than the open-source Qwen2.5-VL-72B (32.00) in the ablation.
- The table lists Qwen2-5-VL-72B in the ablation but the main method is named AWM-Nav, and the 72B row duplicates the main-result numbers exactly.
Limitations
- Relies on pretrained large language models, leading to navigation errors from instruction misjudgment.
- Involuntary mode switching in the dual-tempo action generator.
- Success is defined by a 1.5 m Euclidean distance rather than field-of-view visibility, which relaxes the task versus DDN.
Experimental setup
Datasets: ProcTHOR
Benchmarks: TP-MDDN
Metrics: SR ISR ISPL STL
Baselines: DDN MO-DDN InstructNav
Builds on: DDN InstructNav WMNav MO-DDN
Closest papers by meaning
- MO-DDN: A Coarse-to-Fine Attribute-based Exploration Agent for Multi-Object Demand-driven Navigation (NeurIPS 2024 · similarity 0.89)
- STRIDER: Navigation via Instruction-Aligned Structural Decision Space Optimization (NeurIPS 2025 · similarity 0.86)
- Versatile Navigation Under Partial Observability via Value-guided Diffusion Policy (CVPR 2024 · similarity 0.85)
- Seeing through Uncertainty: Robust Task-Oriented Optimization in Visual Navigation (NeurIPS 2025 · similarity 0.85)
- Hierarchical Semantic-Augmented Navigation: Optimal Transport and Graph-Driven Reasoning for Vision-Language Navigation (NeurIPS 2025 · similarity 0.85)
- Scene Map-based Prompt Tuning for Navigation Instruction Generation (CVPR 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.