Response to Reviewer 7Quy

We sincerely thank the reviewer for their thorough review and insightful feedback.

Response to Weakness 1: Instruction Generation and BreakLLM's Role

• We appreciate the reviewer's feedback regarding our instruction set design. The core purpose of BreakLLM is to decompose abstract human instructions into executable subtasks. Our current instructions are structured to ensure that each subtask's requirements are clear and unambiguous, which is essential for accurately evaluating the agent's understanding and execution, and helps prevent failures caused by ambiguous descriptions. We strongly agree that enhancing realism is important and are considering adopting a more natural instruction format. However, these instructions will still only convey basic needs and preferences without explicitly naming objects, in order to preserve the core challenge of the task—consistent with the instruction set used in our main article. Thank you very much for your suggestion.
• An example of more colloquial instructions is as follows: "Before the guests arrive, please organize the sports equipment using some storage tools. Also, I'd like to set up a small beverage station, which requires preparing some drinkware, and please do a quick clean of the bathrooms with hygiene products to ensure everything looks good." We believe BreakLLM remains effective under such instructions.
• The uniform initialization of all subtask completion statuses to False in BreakLLM is intentional, as it accurately reflects the logical starting point: no subtasks have been completed immediately after parsing a long-horizon instruction. This procedure provides a clear initial value for subsequent modules (e.g. StatusLLM) to accurately monitor and update progress.

Response to Weakness 4: Design for Handling Consecutive Failures

We thank the reviewer for the insightful suggestion. While our current approach effectively mitigates immediate issues, we acknowledge that it can be further improved and extended with a more in-depth analysis.

1. LLM Base Impact on Planning: The choice of large language model (LLM) significantly influences planning quality. In our experiments, Qwen3-32B generated fewer repeated object predictions than Qwen2.5-VL-7B under identical conditions—using the same segmentation module and prompts. This suggests that the base model’s reasoning capabilities directly affect planning reliability, particularly in processing complex multimodal interaction histories. Our findings underscore the importance of advancing LLM backbone performance for instruction-following navigation tasks.

2. Cause of Consecutive Failures and Our Design Inspiration: A typical cause of repeated failures arises when targeting large objects (e.g., a sofa). The LocateLLM predicts the object's center, and the Waypoint Predictor generates a trajectory ending within it. If the agent fails to reach this point, the LLM may re-select the same object as the next target, creating a failure loop. Since precise 3D object sizing remains challenging, we instead leverage auxiliary information to break this cycle—demonstrating it to be an effective and concise solution.

3. Future Work Toward a More Elegant Solution: We agree that a more elegant approach is desirable. Our future work will explore adaptive trajectory planning using object size estimates derived from the 2D map, generating approach points outside the object's boundary based on its dimensions. This aims to enable robust navigation without relying on ad hoc fixes.

Response to Questions

Response to Questions 1:

• Thank you for the feedback. We apologize for any confusion caused by the wording in L206 and will correct it. To clarify, both the slow-tempo and fast-tempo policies use discrete actions in our simulation environment, where fundamental actions like MoveAhead, RotateRight, and Lookup are inherently discrete. The difference lies in their execution: (1) the fast-tempo policy (Lines 125–126) executes individual discrete actions—e.g., a single 0.25m MoveAhead or 30° rotation step; (2) the slow-tempo policy (L206) refers to the outcome of a sequence of discrete actions, where a planned trajectory is broken into waypoints, each reached via multiple discrete steps. Slow-tempo movements use 0.25m steps and 45° rotations, chosen to balance precision and efficiency.

Response to Questions 2:

• Thank you for pointing out the potential source of confusion regarding Figure 2 and line 235. To clarify: (1) The adaptive error corrector is triggered when the fast-tempo process fails to execute a specific discrete action toward a given waypoint. Therefore, the environmental feedback it receives corresponds to discrete actions such as MoveAhead. (2) Furthermore, when the adaptive error corrector encounters issues like collisions, it invokes the waypoint predictor to replan the affordance map and fine-tunes the action interval design. Since this mechanism is closely tied to the planning strategy used in the fast-tempo process, it is logically situated within the fast-tempo framework.

Response to Questions 3:

• We appreciate this important question. Our agent operates in a closed-loop manner: after each discrete action, it receives immediate environmental feedback (e.g., success or collision or boundary detection). If an action fails, the system quickly adjusts the strategy, enabling robust navigation—this is key to the effectiveness of components like the Adaptive Error Corrector. The same feedback is available to other researchers using the environment and the difference lies in whether they choose to leverage it effectively.

Response to Questions 4: Details on Re-planning Mechanism

We appreciate the opportunity to elaborate on our re-planning mechanism:

1. Re-planning Triggers: Failures typically arise from collisions during forward movement (e.g., in narrow passages, with table corners, or partially open doors) or boundary violations. When detected, re-planning is initiated to navigate around the obstacle.

2. Need for Refined Re-planning: We have attempted to input current observation images and context to directly prompt the LLM to generate a sequence of actions to resolve the current obstruction situation. However, experiments showed that renowned large models such as DeepSeek, GPT-4o, and Qwen2.5VL-72B were unable to generate effective action sequences, possibly due to a lack of experience in navigation trajectory planning within simulated environments. This motivated our geometric re-planning strategy.

3. Refined Re-planning Approach. The approach consists of three key components: First, a new affordance graph is generated for the agent's immediate surroundings, enabling a local update of the environment representation. Second, the central coordinate of the original target object remains unchanged, ensuring continuity in the goal. The core innovation lies in the variable sampling interval: when re-planning the trajectory, sampling points are placed at smaller intervals near the beginning (head) of the trajectory, while maintaining the original sampling interval for the latter portion. Experimental results demonstrate that this method effectively helps agents navigate around obstacles and significantly improves the overall task success rate.

4. Validation & Example: To better understand the Adaptive Error Corrector section (illustrated in the lower right corner of Figure 3 in the main text), please refer to the animation task in the far left panel of 'simulation-robot-demo.mp4' (included in the attachments) at 0:01:08. The task instruction is "Organize sports equipment storage, set up a beverage station with drinkware, and maintain bathroom cleanliness with hygiene products". At t=34 seconds, the agent detects the target object in the bathroom within the panoramic view (note that not all panoramic views observed by the agent are shown in the animation). The agent initially plans a trajectory toward the target, but this path is blocked by a closed door. Upon detecting the collision, the system recognizes the action failure and triggers a replanning process. From the agent's current position, a new trajectory to the target object is generated. Thanks to a very small sampling interval for the head trajectory, the agent is able to first execute a turn maneuver, successfully navigating around the obstacle.

Clarity Questions from Weaknesses Section

Response to Weaknesses 2:

Global Records (Long-Term Memory) stores visited information: a cumulative map of detected objects and their 2D poses, and a history of planned targets with their execution outcomes (success/failure), enabling progress tracking. Local Updates (Short-Term Memory) holds transient data for immediate decisions: local 3D point clouds extracted from the current panoramic view (8 perspectives), environmental feedback (e.g., from repeated execution failures), and the current subtask status.

Response to Weaknesses 3:

We sincerely thank the reviewer for pointing out this issue and we will revise this in the later version as follows: Equation (2), which formalizes the process described in Lines 185-194, will be explicitly referenced at that location. For Equation (1), corresponding to Lines 175-177, we will clearly define all variables correctly within the equation: $H_t$ (Historical Exploration Information, revised to $T_t$), $M_t$ (object memory bank, revised to $OM_t$), $State_t$ (current subtask list, revised to $Sub_{status}$), and $d_{sub}$ (subtask details, revised to $Subtasks$). These revisions will ensure that all symbols are properly introduced and the equations are fully understandable.

Response to Reviewer gPWR

We sincerely thank the reviewer for their thorough review and valuable feedback.

Response to Weakness 1 & Question 1: Task Preferences vs. Subtask Criteria

• We strongly agree that a clearer differentiation is crucial. Our "task preferences" address a key gap in prior demand-driven navigation (DDN) datasets, moving beyond object recognition to model abstract user intent.

• Motivation for Task Preferences: Early DDN tasks often featured very simple commands like "I want to sleep". Such instructions are highly ambiguous, as "sleeping" could involve a sofa, a bed, or even resting on a desk, which makes inferring intent and target objects difficult. We propose that each subtask should have an implied or explicit preferred task preference to guide agents.

• Differentiating "Object Preferences" from "Task Preferences":

(1) For a macroscopic instruction like "organize the living space", user preferences can be categorized. "Object preferences" would involve specifying particular objects (e.g., "use a broom", "find a storage box"). However, introducing explicit object information can easily simplify the task to object navigation.

(2) In contrast, our "task preferences" specify the preference of task type, and we ensure that the name of the target object will not be included in the preference. Examples include "achieve cleaning tools", "prepare decorative items," or "find storage tools," as highlighted in lines 23-25 of our main text. This design ensures that the agent must infer the most suitable objects based on the abstract task preference, rather than relying on explicit object names.

(3) Therefore, our "task preferences" are not merely detailed task criteria. The clarity of this requirement is important for correctly executing an abstract basic requirement

Response to Weakness 2: Reliance on Existing Components

• We appreciate the reviewer for the thoughtful feedback. We clarify that our core contribution goes beyond combining existing components. It lies in: (1) identifying previously overlooked limitations of current methods in multi-demand navigation, and (2) introducing novel technical designs to effectively address these challenges, enabling robust and efficient performance.

• Challenges and our contributions:

(1) Map construction: Previous work directly aggregates all encountered point clouds, which performs well over short distances but incurs high computational overhead and slows down reasoning in long-range exploration. To address this, we propose the Multi-dimensional cumulative Semantic Map (MASMap), a novel contribution designed to efficiently manage spatial semantic memory for long-horizon navigation. Its core innovation lies in sophisticated fusion strategies that leverage Ram-Grounded-SAM for object detection, Euclidean distance for spatial relationships, and IoU for object overlap to extract 3D object point clouds from only local 8-view images. More importantly, our global 2D semantic map explicitly tracks and associates observations across multiple views, effectively preventing redundant mapping and duplicate writes for the same object, thereby ensuring a compact, consistent, and efficient representation essential for scalability in large-scale environments.

(2) Dynamic Objective Decision-Making: Current large-model-based demand-driven navigation methods primarily rely on map-point-driven action generation, making them prone to local repetitive planning and lacking dynamic planning awareness for multiple, evolving tasks. To address these challenges, we propose the Autonomous Decision World Model System (AWMSystem), which represents a significant advance in enabling dynamic decision-making without requiring additional training. It integrates an effective chain-of-thought mechanism for object and pose prediction, an adaptive error corrector based on real-time environmental perception to ensure robustness, and a multi-modal large-model-aided task state tracker for more accurate progress monitoring and goal management.

(3) An Novel Evaluation Task for Multi-Demand-Driven Navigation, as important as explained in Response to Weakness 1 & Question 1. Our extensive experiments on this benchmark show that our method achieves state-of-the-art (SOTA) performance in both overall task success rate and subtask completion rate.

(4) Balancing Zero-Shot Transfer Ability and Efficient Reasoning. As shown in Table 2 (our ablation study), the design choices and individual contributions of the components have a significant impact on navigation performance, demonstrating the essentiality of each component.

Response to Weakness 3 & Question 3: Computational Efficiency and Failure Cases.

Computational Efficiency

• We agree that a deeper discussion on computational efficiency and failure cases is critical for real-world applicability. The average execution time for a complete task is approximately 17.41 minutes(nearly 39.75 steps).
• The average time consumption of each module for one task is as follows. BreakLLM :2.89 seconds. Fast-tempo Execution: 67.49 seconds (nearly 32 steps, averaging 2.02 seconds per step). Slow-tempo Execution: 870.74 seconds (nearly steps, averaging 1.81 minutes per step). StatusLLM: 93.93 seconds (averaging 4.68 seconds per call). Adaptive Error Corrector: 12.64 seconds.
• Smaller modules in slow-tempo execution is as follows. Real-time Mapping: 93.68 seconds. LocateLLM: 5.06 seconds. Affordance: 4.13 seconds. Waypoint Predictor: 0.75 seconds. Trajectory Execution: 5.24 seconds.
• It takes about 0.39 seconds for a robot to move or rotate once. This detailed breakdown demonstrates how our dual-tempo design balances robustness with efficiency.

Analysis of Failure Cases

We observed recurring failures related to accuracy and object placement, yet our method still remains viable for real-world use, as shown in our real-robot experiment (see attachment). (1) In cluttered scenes, limited spatial reasoning may cause the LLM to treat nearby objects as separate, leading to redundant visits. However, this occurs significantly less often than with prior map-point-driven methods. We can attempt to enhance the coordinate perception awareness of large models. (2) Small Object Misses: Very small items such as clips or pencils may be missed despite RAM-Grounded-SAM's robustness. Future work could explore zoom-in detection. (3) Hidden Targets: Hidden Targets: If an occluded target is deemed absent, the agent may skip the area, causing failure. This suggests a need for active reasoning, such as checking cabinets, through knowledge bases or uncertainty-driven exploration.

Response to Other Comments & Question 5: Component Selection.

We acknowledge the reviewer's comment regarding adding expanded discussion.

Justification for Component Selection and Integration:

Our system is modularly designed, comprising map construction, command decomposition, and dual-tempo action generating modules. Each component was carefully selected and designed to address specific limitations of prior work in the context of multi-demand long-range navigation:

- component :

1. Map Construction: Problem: Methods like InstructNav store all point clouds, causing slow reasoning (~1 hour/task). Our Fix: We merge 3D points using panoramic views and anti-repetition compression, enabling scalable, fast mapping.

2. Target Object Prediction (LocateLLM): Previous work relies on map-point-driven target prediction, often leading to redundant exploration within the same area. In contrast, our LocateLLM performs object-level prediction, significantly reducing the likelihood of repeated visits. Moreover, we innovatively predicts the target coordinates directly, providing precise spatial input that meets the requirements for subsequent trajectory planning.

3. Adaptive Error Corrector: Despite effective path planning, collisions can still occur in tight spaces. We attempted to use multimodal LLMs for correction, but the action sequences they generated failed to smoothly navigate around obstacles due to a lack of simulation-based planning experience. In contrast, our adaptive corrector leverages real-time environmental feedback to autonomously resolve collisions and boundary issues, significantly improving navigation robustness.

- Ablation Study Insights (Table 2):

1. Segmentation: Grounded-SAM outperformed GLEE (used in DDN/InstructNav), especially in detecting fine details (e.g., cup on toilet)

2. State Monitoring: Using LocateLLM to simultaneously perform status prediction failed, likely because the multitasking prompt diluted its focus and impaired judgment accuracy. In contrast, our hybrid approach, where the expert module triggers "Done" and the LVLM independently confirms completion, significantly improved success rates.

3. LLM Comparison: Qwen2.5-VL-7B exhibited repeated targets, while GPT-4o switched subtasks erratically. In contrast Qwen3-32B demonstrated a more reasonable rhythm in subtask switching.

Response to Question 4: Traditional Planning Algorithms

Traditional planning algorithms (e.g., PDDL) are unsuitable for replacing LocateLLM because they: (1) focus on temporal or priority constraints, while our method emphasizes interpreting multimodal inputs to fulfill abstract instructions; (2) require discrete, predefined state spaces that conflict with our unseen, continuous 3D environments; (3) struggle with dynamic, abstract sub-goals and adaptive task prioritization based on evolving context; and (4) lack robust real-time error correction from continuous sensory feedback. Unlike approaches such as AutoTAMP that use LLMs for output correction, our method employs LLMs for core chain-of-thought reasoning and multimodal understanding.

Details:

Response to Question 2. Subtasks are evaluated as binary outcomes, either successful or failed.

Dear Reviewer,

I hope this message finds you well. As the discussion period is nearing its end with less than two days remaining, I want to kindly check whether we have addressed all your concerns satisfactorily. If there are any further points or suggestions you feel we could clarify or improve, we would be truly grateful for your guidance. Your insights are invaluable to us, and we are eager to address any remaining issues to improve our work.

We deeply appreciate the time, care, and thoughtful effort you have devoted to reviewing our manuscript. Thank you once again for your invaluable contribution.

Best regards,

Thank you for your detailed and thoughtful responses. You’ve addressed my concerns regarding the conceptual differentiation between task preferences and detailed task criteria, provided valuable insights into your component selection and integration choices, and offered an extensive breakdown of computational efficiency and an analysis of failure cases. These clarifications enhance my understanding of your contributions.

To further strengthen the manuscript, I suggest the following minor points for additional clarification:

• The provided computational time breakdown is valuable. Contextualizing these results by directly comparing your method’s efficiency against baselines could further highlight its practical strengths and justify the real-world applicability more convincingly.

• Regarding traditional planning algorithms, your justification effectively notes their limitations in handling multimodal inputs and dynamic adaptability. However, a more explicit comparative reference or supporting citation to prior work demonstrating these limitations would further solidify your argument and provide clearer guidance to the reader.

Overall, your responses have improved my assessment of the paper, and I encourage these minor additions to further enhance clarity and rigor.

Response to Reviewer J2bP

We sincerely thank the reviewer for their thorough review and valuable feedback.

Response to Weakness 1: Justification for Demand-Driven vs. Object-Driven Navigation.

We'd like to clarify that our approach does not directly decode instructions into fixed target objects at the outset. Instead, our demand-driven framework addresses several critical challenges that go beyond the scope of traditional object-goal navigation:

• Dynamic and Abstract Target Selection: Unlike object-goal navigation, where a specific object is the fixed target, our system first identifies objects and their positions from the scene. Then, a large language model (LLM) dynamically determines which sub-task to execute currently and which object to select as the target based on the abstract instruction and the evolving environment. This process also critically involves ensuring that previously selected objects are not duplicated. This dynamic, context-aware decision-making for abstract instructions is a core challenge of demand-driven navigation that object-goal navigation cannot directly resolve.

• Robustness to Object Recognition Uncertainty: In real-world and simulated environments, object recognition is not 100% accurate due to factors like lighting, shadows, and color distortion. Even with advanced segmentation models like RAM-Grounded-SAM, the perceived names of the same object from different perspectives can be inconsistent or inaccurate. Our system accounts for this by requiring the agent to explore and verify whether a recognized object truly meets the abstract demand, rather than blindly navigating to a pre-defined object.

• Real-time Adaptation to Unknown Scenes: We cannot determine the ideal target object for abstract requirements in an unknown scene from the very beginning. Our system continuously records newly detected objects in real-time and feeds them as optional targets to LocateLLM. This means that for the same abstract demand, the LLM dynamically adjusts its target object selection based on newly recognized objects during exploration, enabling adaptive behavior in unseen environments.

Therefore, our demand-driven navigation method is essential, which are fundamentally distinct from and more complex than simple object-goal navigation.

Response to Weakness 2: Methodological Novelty.

We thank the reviewer for their detailed assessment. We respectfully clarify that our primary contribution is not merely an engineering achievement. Instead, it lies in identifying specific, unresolved problems within the context of the novel multi-demand-driven navigation task and proposing innovative technical designs to overcome these challenges. This constitutes significant methodological novelty. Our key technical contributions are:

• Efficient and Scalable Spatial Memory Constructing in long-horizon navigation. Previous work directly aggregates all encountered point clouds, which performs well over short distances but incurs high computational overhead and slows down reasoning in long-range exploration. To address this, we propose the Multi-dimensional cumulative Semantic Map (MASMap), a novel contribution designed to efficiently manage spatial semantic memory for long-horizon navigation. Its core innovation lies in sophisticated fusion strategies that leverage Ram-Grounded-SAM for object detection, Euclidean distance for spatial relationships, and IoU for object overlap to extract 3D object point clouds from only local 8-view images. More importantly, our global 2D semantic map explicitly tracks and associates observations across multiple views, effectively preventing redundant mapping and duplicate writes for the same object, thereby ensuring a compact, consistent, and efficient representation essential for scalability in large-scale environments.

• Dynamic Objective Decision-Making for Multiple Needs: Current large-model-based demand-driven navigation methods primarily rely on map-point-driven action generation, making them prone to local repetitive planning and lacking dynamic planning awareness for multiple, evolving tasks. To address these challenges, we propose the Autonomous Decision World Model System (AWMSystem), which represents a significant advance in enabling dynamic decision-making without requiring additional training. It integrates an effective chain-of-thought mechanism for object and pose prediction, an interaction mechanism based on environmental real-time perception, and a multi-modal large-model-aided task state tracker for more accurate progress monitoring and goal management.

• High-Precision Motion Generation for Robust Navigation: Nowadays, demand-driven action generating methods include using large models to predict the the current one action. However, agents are prone to being attracted by multiple objects in the scene and deviating from the previously planned path, causing decision-making confusion. Some methods approach the detected related objects, but they are often difficult to apply in unseen scenes due to urgent problems such as being unable to pass through narrow doors and getting stuck at table corners. To fill this gap and address issues such as door frame obstruction and exceeding boundaries without additional training, we propose a waypoint predictor targeting the center coordinates of objects, with multi-size collaborative correspondence (agent step size, object size, waypoint sampling interval), and then use an adaptive action corrector based on environmental feedback. Ultimately, in the multi-demand navigation task, we achieved a trade-off between success rate and reasoning performance by using the Dual-Tempo action generation method with efficient approximation of the target object and fine expert detection.

Questions

Response to Question 1: Distance Metric.

• Thank you for the question. To clarify, we define successful sub-task completion using the 2D Euclidean distance between the agent and the center of the target object, with a success threshold of 1.5 units. While benchmarks such as Habitat's HM3D ObjectNav require visual observation of the target object, we do not explicitly verify such observation during evaluation. This choice was made to reduce complexity in the context of multi-demand driven navigation, as discussed in the main text (lines 118–122). A consistent evaluation protocol was applied across all baseline experiments. Furthermore, extensive visual comparisons demonstrate that our method achieves strong performance in accurate target exploration, effective obstacle avoidance, and the resolution of complex multi-demand tasks.

Response to Question 2: LLM Coupling with Map Size / Performance Drop

• We confirm that the size of the semantic map does not directly affect the number of LLM prompts or lead to a decline in LLM performance. This is because our prompt generation is object-based rather than map-based. we detect objects within the agent’s current field of view, compute their 3D coordinates in the camera coordinate system, and then transform them into the real-world coordinate system based on the current camera pose. Only the names and center coordinates of these detected objects are included in the prompt, meaning the number of prompts depends solely on the visible objects at any given moment, not on the overall map size. Furthermore, our 2D semantic grid map—constructed by extracting and merging 3D point clouds from panoramic RGB-D images (captured from 8 perspectives), projecting them onto a 2D plane, and compressing them into a pixel grid—serves primarily as lightweight spatial memory, increasing storage by only kilobytes to megabytes. Crucially, this map is not fed directly into the LLM as input, and therefore does not contribute to prompt length or impact LLM performance.

Response to Question 3: Look Up/Down in Real-World Experiments

• We acknowledge that, although our method supports "look up" and "look down" actions, these are not explicitly included in the real-world experiments. This limitation arises from two main factors: first, the camera on our real robot is mounted along the arc of the mechanical arm, which imposes mechanical constraints that make precise control of vertical gaze directions challenging. Second, the tasks in our real-world scenarios primarily involve common daily-life objects, which can be effectively observed from the robot’s standard horizontal (forward-facing) camera perspective. As such, the absence of explicit vertical viewpoint changes does not significantly hinder the robot’s ability to perceive relevant objects in the environment.

Response to Reviewer DyKS

We sincerely thank the reviewer for the insightful and constructive feedback.

Response to Weakness 1: Motivation for Merging Unrelated Demands.

• Thank you for the valuable feedback. We designed this multi-demand benchmark to evaluate and advance embodied AI in multi-task, long-range navigation. Our design considerations fall into two main aspects. First, we aim to enhance agents' multi-task navigation capabilities, as current single-task navigation accuracy is still limited, and multi-task scenarios tend to accumulate errors. Second, given the limited variety of object types in existing simulation environments—which constrains the creation of multiple related sub-tasks under a single overarching goal—we have instead constructed unrelated sub-tasks. This approach is a pioneering exploration in addressing multi-unrelated-demand-driven navigation challenges.

Here are more specific explanations:

• Addressing Cumulative Error in Sequential Execution: Current single-demand navigation methods often exhibit low success rates. For example, even a strong baseline like InstructNav achieves only 30.0% success on the DDN dataset for single demand. When multiple unrelated demands are executed sequentially, the errors accumulate, leading to a significantly lower overall success rate.

• Bridging the Gap in Multi-Demand Paradigms: We identify a notable gap in existing demand-driven navigation methods for their limited performance in handling multi-demand tasks. While some approaches, such as MO-DDN, suggest potential for multi-step, related requirements (as indicated in its teaser), their datasets and prompts typically focus on fulfilling single preference with high precision. In contrast, our work proposes a novel framework specifically designed to improve success rates on instructions with several high-level demands that do not explicitly name the required objects.

• Comparison with experiment that executes single tasks multiple times. Verifiably, we also proved through experiments that the success rate of the DDN method in successively and separately executing each sub-task is much lower than that of our method. Their results are SR=16% and ISR=47.74%. This indicates that the method of splitting instructions into several individual demand instructions for sequential execution with the aid of large models does not lead to a significant increase in the success rate of multi-demand instruction navigation. It is still very necessary to propose innovative methods to enable agents to efficiently solve multi-demand long-term tasks.

Response to Weakness 2: Clarification on MO-DDN Description.

• We sincerely appreciate the reviewer's valuable clarification regarding the nature of MO-DDN. We fully agree with the reviewer's understanding that MO-DDN is indeed a demand-driven navigation method. To clarify our original statement, the "emphasizes object names" we mentioned in Line 32 of the main text actually refers to the Task Dataset Generation prompt and the example it generates, detailed in Section A.1.1 of the MO-DDN attachment. In that section, the preference requirements for generating raw tasks in their prompt contain "Detail any additional preferences that refine the object selection", which may introduce actual object information and simplify the instructions. For instance, just as their examples on page 18 prompts like "I hope the bed is comfortable enough" explicitly include an object name ("bed") alongside a preference. In contrast, our evaluation set is designed with a stricter requirement: instructions must not contain the name of the target object, and instead express abstract requirements and include task-preferenced preference. It should be noted that the toys in the subtask "Organize a cozy play area with toys" in Figure 2 of our main text are not our target objects. This design ensures the agent to infer the most suitable objects or locations based on abstract demands and highlights one of key difference compared to MO-DDN.

• We will revise lines 33-35 in the manuscript to ensure this distinction is more accurately and clearly communicated, preventing any misinterpretations. Thank you very much for your suggestion.

Response to Weakness 3: Novelty and Differences from Related Work.

Our novelty lies not merely in introducing entirely new concepts, but in the innovative integration, significant refinement, and novel application of these concepts within a holistic framework.

- Differences from Related Work

1. Comparison with InstructNav

(a) Long-range memory banks. InstructNav directly aggregates all encountered point clouds. This approach works well for short distances, as shown by its average trajectory length of 4.44, but causes high computational overhead and slows down reasoning in long-range exploration. Our approach innovatively proposes the Multi-dimensional cumulative Semantic Map (MASMap) construction method. This method creates a highly efficient 2D semantic graph by intelligently merging 3D point clouds from a local panoramic perspective, crucially avoiding duplicate perspective records in the global scope. This design significantly enhances the efficiency and scalability of semantic mapping for truly long-range navigation, a critical distinction from InstructNav's more direct aggregation.

(b) Affordance maps. Instructnav's affordance graph integrates multiple heterogeneous affordance values (e.g., planning direction, action feasibility, landmarks). This fusion process may distort the spatial relevance, causing the highest-scoring location to deviate from the vicinity of the most semantically relevant object in the current context. Furthermore, since instructnav takes the points with the largest afforsibility value as the target, given the high density of points, this method often results in consecutive points being chosen from the same local region, trapping the agent in repetitive, small-scale movements without effective exploration. Therefore, we propose a trajectory planning method that uses an affordance map for precise pathfinding directly to the target object's location, combined with an adaptive error corrector to enhance planning effectiveness and avoid local oscillations.

2. Explanation of Dual-Tempo Action Generation

Thank you very much for your insights. Our Dual-Tempo Action Generator is specifically designed to address the inherent trade-off between the meticulous, but slow, planning capabilities of large models and the rapid execution of expert strategies in the context of long-range embodied navigation. We explicitly highlight this distinction by demonstrating a stark difference in planning times: the large model planning takes 1.81 minutes in one step, whereas the fast-process expert strategy executes in just 2.02 seconds in one step.

- Core Novelties of Our Paper

• Multi-dimensional cumulative Semantic Map (MASMap): This is a novel contribution specifically designed to tackle the challenge of efficiently managing spatial semantic memory in long-range navigation. Its core innovation lies in its sophisticated fusion strategies, leveraging Ram-Grounded-SAM for object detection, Euclidean distance for spatial relationships, and IoU for object overlap, to extract 3D point clouds of objects from only local 8-view images. More importantly, our global 2D semantic map explicitly tracks and associates observations across multiple views, effectively preventing redundant mapping and duplicate write operations for the same object, ensuring a compact, consistent, and efficient semantic representation crucial for scalability in vast environments.

• Autonomous Decision World Model System (AWMSystem): This system represents a significant step towards dynamic objective decision-making in multi-demand tasks without requiring additional training, which integrates an effective chain of thought for object/pose prediction, an adaptive error corrector based on dynamic environmental perception for real-time robustness, and a multi-modal large model-aided task state tracker for more accurate progress and goal management.

• An Novel Evaluation Task for Multi-Demand-Driven Navigation. This benchmark introduces subtasks that deliberately omit explicit object names, express abstract requirements, and incorporate task-type preferences, challenging agents to move beyond simple object retrieval and demonstrate genuine understanding of complex, real-world demands. Our extensive experiments on this benchmark show that our method achieves state-of-the-art (SOTA) performance in both overall task success rate and subtask completion rate.

Response to Question 1

• We are grateful for the reviewer's valuable suggestions. In our baselines comparison, the original DDN method does not require a segmentation model, and MO-DDN uses the same segmentation model as us. As for instructnav, to make a fairer comparison, we conducted a set of experimental results after using the same segmentation model as ours by using RAM-Grounded-SAM, and InstructNav achieved a 14% SR (Success Rate), 3.34 STL (Success weighted by Trajectory Length), 39.39% ISR (Independent Success Rate), and 20.80% ISPL (Independent Success weighted by path Length). These results highlight that even with identical segmentation capabilities, InstructNav's performance remains significantly inferior to ours, demonstrating our framework's superior efficacy and robustness in complex navigation tasks.