Technical Summary: EvolveNav

Problem Statement

Zero-Shot Object-Goal Navigation (ZS-OGN) requires embodied agents to locate specific target objects in unseen environments without prior training or fine-tuning. While recent approaches leverage Foundation Models (Large Language Models and Vision-Language Models) as semantic planners, they suffer from two critical limitations:

1. Static Priors: These models rely on fixed, general semantic knowledge that cannot adapt to real-time feedback, leading to repeated errors (e.g., being consistently misled by visually similar objects).
2. High Trial-and-Error Costs: Existing methods often rely on post-hoc correction, detecting errors only after an action is executed. In environments with limited step budgets, this reactive approach results in inefficient exploration and wasted steps.

The core challenge addressed is how to achieve both online adaptation and action anticipation under strict test-time constraints without updating model parameters.

Methodology: EvolveNav

The authors propose EvolveNav, a training-free framework that transforms navigation into a continuous self-improvement process. The system operates through two tightly coupled components:

1. Self-Evolving Agentic Rule Memory

Instead of updating neural weights, the agent accumulates actionable knowledge in a dynamic rule memory bank ($\mathcal{M}$).

• Rule Extraction: After each episode (success or failure), the system analyzes the trajectory to identify "key steps" that determined the outcome. Using a Semantic-driven Credit Assignment mechanism, it calculates the marginal gain in semantic similarity between the observed scene and the target at each step. These gains are used to assign credit scores, which are then distilled into generalizable navigation rules (e.g., "Avoid areas with no clear indicators").
• Dynamic Management (UCB): To manage the growing rule bank, the framework employs an Upper Confidence Bound (UCB) strategy. The utility of each rule is updated iteratively based on historical success and recent feedback. The UCB score balances the exploitation of proven rules with the exploration of new ones, ensuring that newly generated rules receive an initial validation phase.

2. Memory-Guided Preflection

To mitigate the cost of physical trial-and-error, EvolveNav introduces a preflection module that operates before action execution.

• Proactive Risk Assessment: At each decision step, the agent identifies candidate frontiers (boundaries between explored and unexplored space). Instead of blindly moving, the agent retrieves the top-$K$ relevant rules from the memory bank using the UCB criterion.
• LLM Reasoning: The LLM integrates these retrieved rules with current local observations (visual objects, spatial layout) to perform a "preflection." It explicitly predicts the failure risks of each candidate frontier (e.g., "Frontier 1 is likely a dead-end based on Rule 2").
• Decision Making: The agent selects the optimal frontier by balancing failure probabilities against exploration rewards, effectively filtering out low-value or high-risk paths before they are traversed.

System Architecture

The framework utilizes a hierarchical design:

• Low-Level: A standard occupancy grid map and a low-level controller handle geometric navigation and obstacle avoidance.
• High-Level: The LLM (Instruction Interpolator) handles rule retrieval, preflection reasoning, and waypoint selection.
• Backbone: The system uses a frozen BLIP-2 for vision and Qwen3-8B for language, with all parameters remaining frozen during deployment.

Key Contributions

1. Self-Evolving Rule Memory: A mechanism to support online experience accumulation without parameter updates. By using a UCB-based strategy to balance semantic relevance and historical success, the agent continuously refines its behavior.
2. Memory-Guided Preflection: A shift from passive, post-hoc correction to proactive risk avoidance. By anticipating potential failures before execution, the module mitigates the physical costs of trial-and-error.
3. State-of-the-Art Performance: Extensive experiments demonstrate that the method significantly outperforms existing zero-shot baselines, achieving a 10.1% improvement in Success Rate (SR) on the MP3D dataset while minimizing redundant exploration.

Experimental Results

The method was evaluated on the HM3D and MP3D benchmarks, comparing against both learning-based and zero-shot state-of-the-art (SOTA) methods.

• Quantitative Performance:

  • HM3D: Achieved an SR of 67.3% and SPL of 33.9%, surpassing the previous best zero-shot method (ASCENT) by 1.9% in SR.
  • MP3D: Achieved an SR of 49.0% and SPL of 19.1%, representing a substantial 4.5% absolute improvement in SR over the strongest zero-shot baseline.
  • The authors note that the performance gap widens in complex, multi-floor environments (MP3D), where static priors typically fail.

• Ablation Studies:

  • Removing Preflection resulted in a 3.5% SR drop on MP3D, validating the importance of proactive risk avoidance.
  • Removing Memory-Evolving (using a static rule set) caused a 4.5% SR drop on MP3D, highlighting the necessity of online adaptation.
  • The framework showed limited sensitivity to the specific scale of the LLM backbone (Qwen2.5-7B vs. Qwen3.5-9B), suggesting the architectural design (memory + preflection) is the primary driver of performance rather than raw model size.

• Qualitative Analysis: Visualizations show that EvolveNav avoids the "convoluted" paths and local traps (e.g., narrow corridors, irrelevant small rooms) that plague baseline methods like ASCENT. The agent successfully bypasses misleading areas (e.g., spa/jacuzzi areas when looking for a potted plant) based on learned heuristics.

Significance and Claims

The paper claims that EvolveNav addresses a critical bottleneck in embodied AI: the inability of foundation models to adapt to real-time feedback in unseen environments without costly retraining. By decoupling "learning" from "parameter updates" and instead focusing on rule extraction and dynamic retrieval, the framework enables agents to:

1. Internalize exploration experiences to prevent the repetition of navigational errors.
2. Shift the paradigm from reactive correction to proactive anticipation, significantly reducing the physical cost of trial-and-error.
3. Achieve robust generalization in complex, multi-floor environments where static semantic priors fail.

The authors position this work as a step toward more efficient, self-improving embodied agents that can operate effectively in the real world without the need for massive datasets or environment-specific fine-tuning.