Technical Summary: CLOSER-VLN

Problem Statement

Vision-Language Navigation (VLN) has advanced significantly with the integration of Large Language Models (LLMs) and Multimodal Large Language Models (MLLMs), enabling agents to follow natural language instructions in unseen environments without task-specific training. However, existing zero-shot VLN methods predominantly rely on an open-loop decision-execution approach. In this paradigm, agents generate a candidate action based on instructions and observations and execute it immediately, lacking explicit pre-execution verification.

This open-loop strategy is particularly fragile in Aerial VLN (e.g., UAV navigation in urban environments). Unlike ground-based scenarios, aerial navigation involves longer flight distances, broader operational coverage, and stricter reliability requirements. Minor errors in intermediate actions can rapidly accumulate, leading to significant trajectory deviations and eventual target loss. Furthermore, while Retrieval-Augmented Generation (RAG) has been applied to embodied agents to supplement knowledge, existing methods typically use retrieval as a general context enhancer before or during reasoning. They rarely couple retrieval with action verification to determine when an action is unreliable or how it should be corrected. Consequently, current methods lack a mechanism to detect and rectify errors before they impact the agent's state.

Methodology: CLOSER and CLOSER-VLN

To address these limitations, the authors propose Closed-loop Self-verified Retrieval-augmented Reasoning (CLOSER), a training-policy-free method that transforms navigation into a bounded, iterative closed-loop process. Based on this, they instantiate CLOSER-VLN, a framework specifically designed for aerial VLN tasks.

Core Workflow

Instead of executing the first predicted action, CLOSER-VLN performs a sequence of Action Reasoning $\rightarrow$ Reliability Verification $\rightarrow$ Targeted Retrieval $\rightarrow$ Action Correction before each environment step. This loop continues until an action is accepted or a maximum iteration limit ($K_{max}$) is reached.

The framework consists of three specialized components:

1. Hierarchical Reasoner:

   • Acts as the action generation module.
   • Takes stage-level sub-instructions, environment descriptions, visual inputs (Schematic Cognitive Maps for navigation/search; raw observations for localization), and retrieved knowledge as input.
   • Generates a candidate action and a rationale.
   • Operates across three task stages: Navigation (approaching target regions), Search (identifying target nodes), and Localization (fine-grained coordinate prediction).

2. Multidimensional Action Verifier:

   • Serves as the closed-loop gatekeeper to assess action reliability.
   • For Navigation and Search: Uses a fine-tuned Qwen3-VL-8B model to evaluate three dimensions:
     • Instruction Consistency: Does the action align with the linguistic command?
     • Visual Consistency: Is the action grounded in the semantic cognitive map?
     • Path Feasibility: Does the action avoid backtracking or flying out of bounds?
   • For Localization: Employs a heuristic geometric check to ensure the predicted target coordinate falls within a valid distance threshold ($\tau_{dist} = 50m$) from the current position.
   • Training Strategy: The verifier is trained via a two-stage process: Supervised Fine-Tuning (SFT) to learn basic logic, followed by Group Relative Policy Optimization (GRPO) with an asymmetric reward function. This reward structure heavily penalizes False Positives (accepting wrong actions) to ensure conservative rejection of unreliable decisions.

3. Verification-Triggered Multimodal Retriever:

   • Unlike conventional RAG, retrieval is not performed unconditionally. It is triggered only when the verifier rejects a candidate action.
   • Upon rejection, the system analyzes the error (e.g., specific failure dimensions) to generate a targeted query.
   • It retrieves relevant exemplars from a stage-aligned external knowledge base (constructed from successful decision fragments in the CityNav training set).
   • The retrieved exemplar, combined with the error analysis, is fed back to the reasoner to refine the decision in the next iteration.

Bounded Inference

To prevent infinite loops, the system sets a maximum iteration count ($K_{max} = 3$). If no action is accepted within this limit, the final candidate from the last iteration is executed as a fallback.

Key Contributions

1. CLOSER Framework: Proposes a closed-loop decision correction method for VLN that introduces reliability verification before execution and triggers targeted retrieval only upon verification failure, addressing the difficulty of error detection in open-loop systems.
2. CLOSER-VLN Implementation: Adapts the CLOSER method to aerial VLN, introducing a bounded closed-loop process to improve reliability in unseen urban scenarios.
3. Specialized Components: Designs a Multidimensional Action Verifier (trained with GRPO for high rejection precision) and a Verification-Triggered Multimodal Retriever. These components collaborate to identify unreliable actions and perform corrections using stage-aligned external exemplars.
4. Empirical Validation: Demonstrates significant performance improvements on the CityNav benchmark, validating the effectiveness of closed-loop reasoning over open-loop baselines.

Experimental Results

Experiments were conducted on the CityNav benchmark (Test Unseen, Validation Unseen, and Validation Seen splits).

• Performance Gains: On the Test Unseen split, CLOSER-VLN achieved a Success Rate (SR) of 32.01% and a Success weighted by Path Length (SPL) of 21.28%.
  • This represents a 23.6% relative improvement in SR and a 32.8% relative improvement in SPL compared to the best-performing zero-shot baseline, GeoNav.
  • It also outperformed supervised baselines (e.g., FlightGPT) and other zero-shot methods (e.g., GPT-4o, Greedy).
• Difficulty Levels: The method showed consistent improvements across Easy, Medium, and Hard difficulty levels, with the most significant gains (53.3% relative SR improvement) observed in Hard tasks, which require long-distance planning and precise landmark grounding.
• Ablation Studies:
  • Removing the verifier reduced SR to 29.24%.
  • Using "Always Simple Retrieval" (retrieving at every step regardless of verification) actually degraded performance (SR 26.96%) due to irrelevant information noise.
  • The full system (Verifier + Adaptive Retrieval) was required to achieve optimal results, confirming the synergistic effect of verification and targeted knowledge supplementation.
• Iteration Analysis: Performance peaked at $K_{max}=3$. Increasing iterations to 4 caused performance to decline due to the introduction of noise from erroneous retrieval queries.

Significance and Claims

The paper claims that CLOSER-VLN effectively addresses the critical limitation of error accumulation in open-loop aerial navigation. By shifting from a "predict-and-execute" paradigm to a "reason-verify-retrieve-correct" loop, the system can:

• Detect and correct unreliable actions before they alter the agent's state.
• Dynamically supplement missing knowledge only when necessary, avoiding the noise of indiscriminate retrieval.
• Achieve robust generalization in unseen urban environments without training a task-specific navigation policy.

The authors emphasize that the multidimensional verifier is crucial for safety-critical applications, as it prevents the agent from blindly executing hallucinated or geometrically invalid actions. The results suggest that closed-loop reasoning significantly enhances the reliability of embodied agents in complex, long-horizon tasks where single-step errors can be catastrophic.

Note: The paper acknowledges limitations, such as the verifier's occasional failure to detect complex geometric boundaries (leading to false positives in acceptance), and suggests future work in optimizing the verifier via incremental fine-tuning and exploring Group Sequence Policy Optimization (GSPO).