Research map / Multi-agent systems and games / Human-AI and goal-conditioned
STRIDER: Navigation via Instruction-Aligned Structural Decision Space Optimization
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 STRIDER, a training-free zero-shot VLN-CE framework that builds skeleton-based structured waypoints from depth and uses VLM-generated task-progress feedback in a closed loop with an LLM planner.
Problem
Zero-shot vision-and-language navigation in continuous environments (VLN-CE) where agents must follow instructions in unseen 3D scenes without training, struggling with execution drift and open-loop action selection over long horizons.
Approach
Extracts a topological skeleton from depth-based navigable regions and uses degree-1 endpoints as layout-constrained waypoints, paired with VLM directional descriptions; a Task-Alignment Regulator compares consecutive observations with a VLM to produce progress feedback that conditions the LLM's waypoint choice at each step.
Key results
- On R2R-CE (100 sampled episodes), STRIDER reaches 35% SR and 30.30 SPL, versus 29% SR for the prior best zero-shot SmartWay, a 20.7% relative SR gain as claimed.
- On RxR-CE, STRIDER achieves 21.2 SR and 9.6 SPL, outperforming zero-shot CA-Nav (19.0 SR, 6.0 SPL).
- Ablation: removing the Task-Alignment Regulator drops SR from 35 to 29 and SPL from 30.30 to 26.11 on R2R-CE.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use STRIDER instead of Open-Nav-GPT4 | zero-shot navigation · R2R-CE | NDTW: 45.79 -> 51.8 | Helped |
| use STRIDER instead of SmartWay | zero-shot navigation · R2R-CE | SR: 29 -> 35 | Helped |
| use STRIDER instead of CA-Nav | zero-shot navigation · RxR-CE | NE: 10.37 -> 11.19 | Hurt |
| use STRIDER instead of CA-Nav | zero-shot navigation · RxR-CE | NDTW: 13.5 -> 30.1 | Helped |
| add Structured Waypoint Generator using degree-1 endpoints (ablation) | STRIDER without SWG · R2R-CE | SR: 24 -> 35 | Helped |
| select only degree > 2 skeleton nodes instead of degree-1 endpoints (ablation) | Structured Waypoint Generator · R2R-CE | SR: 35 -> 28 | Hurt |
| select nodes with degree ≠ 2 instead of degree-1 endpoints (ablation) | Structured Waypoint Generator · R2R-CE | SR: 35 -> 33 | Hurt |
| add Task-Alignment Regulator (ablation) | STRIDER without TAR · R2R-CE | SR: 29 -> 35 | Helped |
| replace Qwen-VL-Max with Qwen2.5-VL-72B (ablation) | STRIDER · R2R-CE | SR: 35 -> 34 | Hurt |
| replace Qwen-VL-Max with Qwen2.5-VL-32B (ablation) | STRIDER · R2R-CE | SR: 35 -> 28 | Hurt |
| replace Qwen-VL-Max with Qwen2.5-VL-7B (ablation) | STRIDER · R2R-CE | SR: 35 -> 24 | Hurt |
| replace BEVBert waypoint prediction with Structured Waypoint Generator (ablation) | BEVBert · R2R-CE | SR: 59 -> 61 | Helped |
Surprising findings
- Degree-1 endpoint waypoints outperform junction-based waypoints, and using only high-degree nodes degrades performance.
- The feedback module slightly increases navigation error (6.77 to 6.91) while improving SR and SPL.
Limitations
- Fixed horizontal camera view misses staircases and vertical elevation changes.
- Relies on large API-based VLMs and LLMs whose perception and reasoning errors propagate; added latency limits real-time use.
- Evaluated only on 100 randomly selected R2R-CE episodes; NE is not best and RxR-CE gains are narrower.
Experimental setup
Datasets: R2R-CE RxR-CE
Benchmarks: R2R-CE RxR-CE
Metrics: NE NDTW OSR SR SPL SDTW TL
Baselines: Open-Nav-GPT4 Open-Nav-Llama3.1 SmartWay DiscussNav-GPT4 CA-Nav A2Nav BEVBert ETPNav HNR
Builds on: Open-Nav BEVBert
Closest papers by meaning
- CompassNav: Steering From Path Imitation to Decision Understanding In Navigation (ICLR 2026 · similarity 0.91)
- SeqWalker: Sequential-Horizon Vision-and-Language Navigation with Hierarchical Planning (AAAI 2026 · similarity 0.90)
- Run, Ruminate, and Regulate: A Dual-process Thinking System for Vision-and-Language Navigation (AAAI 2026 · similarity 0.90)
- Scene Map-based Prompt Tuning for Navigation Instruction Generation (CVPR 2025 · similarity 0.89)
- NavGPT: Explicit Reasoning in Vision-and-Language Navigation with Large Language Models (AAAI 2024 · similarity 0.88)
- Hierarchical Semantic-Augmented Navigation: Optimal Transport and Graph-Driven Reasoning for Vision-Language Navigation (NeurIPS 2025 · similarity 0.88)
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.