Pith. sign in

REVIEW 4 major objections 6 minor 28 references

SemNav: A Model-Based Planner for Zero-Shot Object Goal Navigation Using Vision-Foundation Models

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read SemNav claims that combining GPT-4o likelihood scoring of frontiers with the LSP expected-cost planner achieves state-of-the-art success-weighted path length in zero-shot object goal navigation.

desk verdict A plausible zero-shot object-nav result that combines GPT-4o frontier scoring with LSP planning; the central SPL claim is believable but not yet reproducible, so the paper deserves peer review with requirements for code and calibration analysis. read the letter →

arxiv 2506.03516 v1 pith:THSNC3HC submitted 2025-06-04 cs.RO cs.AI

Arnab Debnath , Gregory J. Stein , Jana Kosecka This is my paper
classification cs.ROcs.AI
keywords objectgoalnavigationzero-shotvision-languagemodelsfrontierexplorationmodel-basedplanningLSPplannerHM3Dsemantic
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

SemNav addresses zero-shot object goal navigation, where a robot must find a target object in an unseen environment without task-specific training. The paper proposes to score candidate exploration frontiers using commonsense visual reasoning from GPT-4o, then select frontiers with a non-greedy, model-based planner (LSP) that estimates the expected navigation cost. On the HM3D validation benchmark, SemNav reports an SPL of 35.9 and SR of 54.9, beating prior zero-shot methods on path efficiency. The authors argue that the combination of richer semantic scoring and long-horizon planning produces more efficient exploration than greedy frontier selection or image-text similarity alone.

What carries the argument

The load-bearing mechanism is the LSP recursive expected-cost equation, $Q(\{m_t, q_t, A_t\}, a_t) = D(m_t, q_t, a_t) + P_S R_S + (1-P_S)(R_E + \max_{a \in A_t \setminus \{a_t\}} Q(\{m_t, q_{a_t}, A_t \setminus \{a_t\}\}, a))$, where $P_S$ is the VLM-assigned probability of finding the target through frontier $a_t$, and $R_S$ and $R_E$ are constant success and exploration costs. The planner selects the frontier minimizing this expected cost, enabling long-horizon reasoning rather than greedy pick-the-highest-score selection.

What would settle it

A direct test would be to rerun SemNav on the same episodes with GPT-4o scores replaced by random or constant values; if SPL remains similar, the scores contribute nothing. Alternatively, measure whether the GPT-4o score assigned to a frontier predicts the actual presence of the target beyond that frontier across many episodes, or repeatedly query GPT-4o on the same image to check whether the floating-point outputs are stable enough to be treated as probabilities.

Watch

Extended reading notes

Core claim

The central discovery is that a vision-language model's single-image likelihood estimate, when inserted into the LSP expected-cost equation as the probability of finding the target beyond a frontier, yields a planner that navigates to object goals more efficiently than all prior zero-shot baselines. SemNav's SPL of 35.9 on HM3D-val surpasses VLFM's 30.4, and the paper attributes this to GPT-4o's ability to reason about plausible object locations even when the target is not visible, combined with LSP's non-myopic cost evaluation. The paper also shows a 1.6 SPL drop when the same scores are used greedily, indicating that the planner itself contributes beyond scoring.

Load-bearing premise

The load-bearing assumption is that the raw floating-point numbers GPT-4o emits for a single egocentric image are meaningful, comparable estimates of the probability of finding the target object beyond each frontier, and that plugging them directly into the LSP cost equation produces a valid expected cost.

Editorial extensions

If this is right

  • If SemNav's reported SPL advantage holds, zero-shot object navigation can be improved through commonsense semantic scoring of frontiers rather than fine-tuning on task-specific data.
  • The non-greedy LSP planner contributes measurably to efficiency, since replacing it with greedy frontier selection drops SPL by 1.6 points.
  • SemNav's SR of 54.9 is close to, though slightly below, intervention-based methods like TriHelper and MFNP, suggesting that pure frontier reasoning can compete with targeted failure-handling modules.
  • The approach's reliance on a pretrained local navigation policy means the semantic reasoning and planning components can transfer to other environments without further training.
  • Correcting the reported simulator issue, which the paper says accounts for about 25 percent of failed episodes, could raise SR further.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Because the GPT-4o scores are used as probabilities without calibration, the LSP formalism may effectively operate as a re-ranking heuristic; testing calibrated or otherwise validated probabilities could either strengthen or weaken the reported advantage.
  • The method could extend to other domains where commonsense spatial reasoning matters, such as outdoor search or aerial navigation, by swapping the VFM prompt and retaining the same cost-based planner.
  • A quantitative analysis of whether GPT-4o's floating-point outputs predict actual target presence beyond frontiers would isolate the source of the SPL gain and suggest a calibration scheme for further improvement.
  • The 25 percent simulator-failure caveat, if resolved, could alter the SR ranking relative to TriHelper and MFNP, changing the comparison story.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. SemNav proposes a zero-shot object-goal navigation system that combines GPT-4o-based frontier scoring with the model-based Learning Over Subgoals Planner (LSP). The agent builds an egocentric geometric and semantic map, extracts frontiers, and at each step queries GPT-4o for a floating-point likelihood of finding the target object in the current viewing direction. These scores are interpreted as probabilities and inserted into the LSP expected-cost recursion in Eq. (1); the frontier with minimal expected cost is selected. The method is evaluated on the HM3D-val benchmark (2000 episodes) in the Habitat simulator, reporting SPL 35.9 and SR 54.9, which the paper claims is state-of-the-art SPL for zero-shot object-goal navigation. The appendix provides qualitative trajectory comparisons and examples of GPT-4o outputs.

Significance. If the empirical claim holds, this is a useful demonstration that VLM-based semantic frontier scoring can be combined with non-myopic planning to improve path efficiency over greedy scoring and over the cosine-similarity scoring used by VLFM. The explicit ablation of SemNav versus SemNav-Greedy is a strength, as it isolates the contribution of the LSP planner. The use of a standard benchmark (HM3D-val) and the qualitative appendix examples also help interpret the behavior. However, the central evidence is a single benchmark table with no repeated runs, no error bars, and no calibration analysis for the VLM scores, so the significance is conditional until those issues are addressed. The paper's contribution would be stronger with code/data release and a robustness analysis of the probabilistic input to Eq. (1).

major comments (4)
  1. [3.2, Eq. (1)] The entire planning recursion treats the GPT-4o output PS as a probability of finding the target object beyond a frontier, but the paper provides no evidence that these outputs are calibrated probabilities, stable across repeated queries, or comparable across images, frontiers, and scenes. The prompt in Sec. 3.2 requests only "a floating point value" with no temperature, no output range, no seed, and no validation. Because Eq. (1) directly multiplies PS with RS and RE, any systematic bias or variance in the scores propagates into frontier ordering. Please add a calibration study (e.g., compare scores with ground-truth object presence in held-out scenes), a stability analysis (repeated queries per image with reported variance), and a perturbation experiment (e.g., monotone transforms or additive noise on PS) to show that the reported SPL does not depend on the precise numeric scale of the VLM outputs.
  2. [5, Table 1] The headline comparison rests on a single reported run with no error bars, seeds, or repeated evaluations, while GPT-4o is non-deterministic and versioned by API. The claimed advantage over SemNav-Greedy is 1.6 SPL points, which could be within run-to-run noise. Please report multiple runs or confidence intervals, specify the GPT-4o version, temperature, and access date, and release code and data to make the result reproducible. If full repeated evaluations are too expensive, a smaller-scale repeated evaluation on a subset with reported variance would still help assess the stability of the ranking.
  3. [5] The paper states that "around 25% of the failed episodes are due to a simulator issue, where the agent successfully reaches an instance of the target object category but the simulator still registers the episode as a failure." This is potentially large enough to change the SR and SPL ranking, but the statement is vague: there is no definition of the simulator issue, no episode counts, no explanation of whether these failures were included or excluded from Table 1, and no evidence that all baselines were affected equally. Please provide a precise accounting and, if possible, re-compute metrics with corrected labels or with a manually validated subset.
  4. [3.3] The constants RS = 3m and RE = 6m are hand-set, and the paper claims "Empirically, we observed that varying these constants has minimal impact on performance" without showing the evidence. Since Eq. (1) depends on the ratio RS/RE, a sensitivity analysis over a grid of these constants is needed to support the claim that the planner's behavior and the reported SPL are insensitive to them. Without this, the headlined SPL result is not robustly established.
minor comments (6)
  1. [Abstract] The sentence "enabling more scalable and adaptable solution" should be "enabling a more scalable and adaptable solution" or "enabling more scalable and adaptable solutions."
  2. [3.1] The value-map update formula appears malformed; it should likely be v_new = (c_curr * v_curr + c_prev * v_prev) / (c_curr + c_prev), but the text has unbalanced parentheses and missing operators.
  3. [5] The phrases "5.5% and 2.4% improvement in SPL and SR" and "1.6% SPL drop" should be expressed as percentage-point differences rather than percentages, since the absolute differences in Table 1 are 5.5 and 2.4 SPL/SR points and 1.6 SPL points.
  4. [3.4] Mobile-SAM is used for segmentation but is not cited in the references; please add the appropriate citation.
  5. [3.2] The prompt should specify the intended output range (e.g., [0,1]) and a definition of "likelihood" to make the VLM query reproducible and to clarify whether GPT-4o is being asked for a probability or a score.
  6. [Table 1] The baseline name "V oroNav" contains an irregular space and should be normalized to "VoroNav".

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: SemNav's SPL claims rest on external benchmark evaluation and an independent planner ablation, not on self-referential derivation.

full rationale

SemNav's central claim is an empirical SPL/SR measurement on the external HM3D validation benchmark (Table 1), reported against independent baselines. This is not a quantity derived from the method's own assumptions. The probability input P_S is produced by GPT-4o as a raw VLM score and is not fitted to the benchmark outcome; the hand-set constants R_S = 3m and R_E = 6m are stated assumptions with an explicit claim that performance is insensitive to them. The LSP planner is the only component with substantial author overlap via citation [15] and related self-citations [7,8], but the paper restates the full LSP recursion in Eq. (1) and tests its contribution through the SemNav-versus-SemNav-Greedy ablation, so the planner's value is supported by an internal controlled comparison rather than by an unverified black-box citation. No equation in the paper reduces to its own input, no fitted parameter is renamed as a prediction, and no known result is disguised as a new derivation. The lack of calibration or determinism testing of GPT-4o scores is a legitimate correctness and robustness concern about whether the SPL advantage transfers, but it is not a circularity: the result is an external measurement, not a tautology. Therefore no specific circular steps are identified.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The method has no invented physical entities. The free parameters are the two LSP cost constants; the other assumptions are standard domain assumptions about detectors, maps, and pretrained policies.

free parameters (2)
  • RS (success cost) = 3 m
    Expected distance to object beyond a frontier if the target is found; chosen by hand in Section 3.3, with no sensitivity analysis shown.
  • RE (exploration cost) = 6 m
    Exploration and return cost if the target is not found beyond the frontier; chosen as twice RS in Section 3.3.
assumptions (4)
  • domain assumption GPT-4o floating-point likelihoods are calibrated probabilities PS usable in LSP's expected cost equation.
    Section 3.2 feeds a single RGB image and prompt to GPT-4o and uses the returned number as the probability of finding the target; no calibration or consistency check is reported.
  • domain assumption The pretrained point-goal navigation policy from VLFM, trained on the HM3D training split, transfers to HM3D-val and is treated as a zero-shot component.
    Section 3.4 reuses the local policy from [20]; this is a learned model trained on the same dataset family, which narrows the zero-shot claim.
  • domain assumption YOLOv7 and Mobile-SAM detections and segmentations are accurate enough for the 6 target categories.
    Section 3.4 uses these off-the-shelf detectors without reporting detection failure rates or missed detections.
  • domain assumption The frontiers extracted from the depth-based partial map are valid navigation targets.
    Section 3.1 assumes the height-threshold point-cloud mapping produces accurate free and obstacle boundaries; no evaluation of map accuracy is given.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SemNav: A Model-Based Planner for Zero-Shot Object Goal Navigation Using Vision-Foundation Models." pith.science (2026). https://pith.science/paper/THSNC3HC

@misc{pith2026250603516,
  author       = {Pith},
  title        = {Pith review of: SemNav: A Model-Based Planner for Zero-Shot Object Goal Navigation Using Vision-Foundation Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/THSNC3HC}},
  note         = {Machine review of arXiv:2506.03516}
}
read the original abstract

Object goal navigation is a fundamental task in embodied AI, where an agent is instructed to locate a target object in an unexplored environment. Traditional learning-based methods rely heavily on large-scale annotated data or require extensive interaction with the environment in a reinforcement learning setting, often failing to generalize to novel environments and limiting scalability. To overcome these challenges, we explore a zero-shot setting where the agent operates without task-specific training, enabling more scalable and adaptable solution. Recent advances in Vision Foundation Models (VFMs) offer powerful capabilities for visual understanding and reasoning, making them ideal for agents to comprehend scenes, identify relevant regions, and infer the likely locations of objects. In this work, we present a zero-shot object goal navigation framework that integrates the perceptual strength of VFMs with a model-based planner that is capable of long-horizon decision making through frontier exploration. We evaluate our approach on the HM3D dataset using the Habitat simulator and demonstrate that our method achieves state-of-the-art performance in terms of success weighted by path length for zero-shot object goal navigation.

Figures

Figures reproduced from arXiv: 2506.03516 by the authors.

Figure 1
Figure 1. Overview of SemNav we build upon the model-based frontier planning strategy, Learning Over Subgoals Planner (LSP) [15], which eval￾uates the expected navigation cost of selecting a frontier given the current belief about the environment. This allows the agent to reason about the long-term consequences of its actions and supports more efficient long-horizon planning. Our contributions are summarized below: • We propo… view at source ↗
Figure 2
Figure 2. BLIP-2 assigns a lower score for the couch from the [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Final navigation paths for the target object bed. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Some example images with their corresponding GPT-4o responses. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: This is an example where we see that SemNav quickly gets out of the room to find a couch than VLFM. Top row: At timestep t=4, [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Row 1: At timestep t = 20, SemNav (with LSP) selects a frontier with a lower score (0.29 in map) to minimize expected travel cost, choosing closer frontiers that guide it toward the target. In contrast, SemNav-Greedy selects the frontier with the highest score and move…
Figure 7
Figure 7. Figure 7: Here, we show a problem with the simulator where the agent successfully reached the target object, but it was still considered [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

28 extracted references · 21 canonical work pages

  1. [1]

    Zero experience required: Plug & play modular transfer learning for semantic visual navigation

    Ziad Al-Halah, Santhosh Kumar Ramakrishnan, and Kristen Grauman. Zero experience required: Plug & play modular transfer learning for semantic visual navigation. InProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17031–17041, 2022. 2

    2022

  2. [2]

    Object goal naviga- tion using goal-oriented semantic exploration.Advances in Neural Information Processing Systems, 33, 2020

    Devendra Singh Chaplot, Dhiraj Prakashchand Gandhi, Ab- hinav Gupta, and Russ R Salakhutdinov. Object goal naviga- tion using goal-oriented semantic exploration.Advances in Neural Information Processing Systems, 33, 2020. 1

    work page 2020

  3. [3]

    Maskclip: Masked self- distillation advances contrastive language-image pretraining

    Xiaoyi Dong, Jianmin Bao, Yinglin Zheng, Ting Zhang, Dongdong Chen, Hao Yang, Ming Zeng, Weiming Zhang, Lu Yuan, Dong Chen, et al. Maskclip: Masked self- distillation advances contrastive language-image pretraining. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10995–11005, 2023. 1

    work page 2023

  4. [4]

    Segment any- thing

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer White- head, Alexander C Berg, Wan-Yen Lo, et al. Segment any- thing. InProceedings of the IEEE/CVF international confer- ence on computer vision, pages 4015–4026, 2023. 1

    2023

  5. [5]

    BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models

    Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. InICML,

    work page

  6. [6]

    Grounded language-image pre-training

    Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jian- wei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10965–10975, 2022. 1

    2022

  7. [7]

    Comparison of Model-Free and Model-Based Learning-Informed Planning for PointGoal Navigation

    Yimeng Li, Arnab Debnath, Gregory J Stein, and Jana Kosecka. Comparison of model-free and model-based learning-informed planning for pointgoal navigation.arXiv preprint arXiv:2212.08801, 2022. 3

    work page Pith review arXiv 2022

  8. [8]

    Learning-augmented model-based planning for vi- sual exploration

    Yimeng Li, Arnab Debnath, Gregory J Stein, and Jana Koˇseck´a. Learning-augmented model-based planning for vi- sual exploration. In2023 IEEE/RSJ International Confer- ence on Intelligent Robots and Systems (IROS), pages 5165–

    work page

Show all 28 references
  1. [9]

    Grounding dino: Marrying dino with grounded pre-training for open-set object detection

    Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. In European Conference on Computer Vision, pages 38–55. Springer...

    2024

  2. [10]

    Zson: Zero-shot object-goal navigation using multimodal goal embeddings.Advances in Neural Information Processing Systems, 35:32340–32352,

    Arjun Majumdar, Gunjan Aggarwal, Bhavika Devnani, Judy Hoffman, and Dhruv Batra. Zson: Zero-shot object-goal navigation using multimodal goal embeddings.Advances in Neural Information Processing Systems, 35:32340–32352,

  3. [11]

    Learning hierarchical relationships for object-goal navigation

    Anwesan Pal, Yiding Qiu, and Henrik Christensen. Learning hierarchical relationships for object-goal navigation. InCon- ference on Robot Learning, pages 517–528. PMLR, 2021. 1

    2021

  4. [12]

    Habitat-Matterport 3D dataset (HM3d): 1000 large- scale 3D environments for embodied AI

    Santhosh Kumar Ramakrishnan, Aaron Gokaslan, Erik Wi- jmans, Oleksandr Maksymets, Alexander Clegg, John M Turner, Eric Undersander, Wojciech Galuba, Andrew West- bury, Angel X Chang, Manolis Savva, Yili Zhao, and Dhruv Batra. Habitat-Matterport 3D dataset (HM3d): 1000 large- s...

    2021

  5. [13]

    PONI: Potential functions for ObjectGoal navigation with interaction-free learning

    Santhosh Kumar Ramakrishnan, Devendra Singh Chap- lot, Ziad Al-Halah, Jitendra Malik, and Kristen Grauman. PONI: Potential functions for ObjectGoal navigation with interaction-free learning. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition,

  6. [14]

    Habitat: A platform for embodied ai research

    Manolis Savva, Abhishek Kadian, Oleksandr Maksymets, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia Liu, Vladlen Koltun, Jitendra Malik, et al. Habitat: A platform for embodied ai research. InProceedings of the IEEE/CVF International Conference on Computer Vision, p...

    2019

  7. [15]

    Learning over subgoals for efficient navigation of structured, unknown environments

    Gregory J Stein, Christopher Bradley, and Nicholas Roy. Learning over subgoals for efficient navigation of structured, unknown environments. InConference on Robot Learning, pages 213–222. PMLR, 2018. 2, 3

    2018

  8. [16]

    Yolov7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors

    Chien-Yao Wang, Alexey Bochkovskiy, and Hong- Yuan Mark Liao. Yolov7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors. InPro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7464–7475, 2023. 3

    2023

  9. [17]

    Exploitation-guided exploration for semantic embodied navigation

    Justin Wasserman, Girish Chowdhary, Abhinav Gupta, and Unnat Jain. Exploitation-guided exploration for semantic embodied navigation. In2024 IEEE International Confer- ence on Robotics and Automation (ICRA), pages 2901–2908. IEEE, 2024. 4

    2024

  10. [18]

    V oronav: voronoi-based zero- shot object navigation with large language model

    Pengying Wu, Yao Mu, Bingxian Wu, Yi Hou, Ji Ma, Shang- hang Zhang, and Chang Liu. V oronav: voronoi-based zero- shot object navigation with large language model. InPro- ceedings of the 41st International Conference on Machine Learning, pages 53737–53775, 2024. 2, 4

    2024

  11. [19]

    Ovrl-v2: A simple state- of-art baseline for imagenav and objectnav.arXiv preprint arXiv:2303.07798, 2023

    Karmesh Yadav, Arjun Majumdar, Ram Ramrakhya, Naoki Yokoyama, Alexei Baevski, Zsolt Kira, Oleksandr Maksymets, and Dhruv Batra. Ovrl-v2: A simple state- of-art baseline for imagenav and objectnav.arXiv preprint arXiv:2303.07798, 2023. 1

    2023 arXiv

  12. [20]

    Vlfm: Vision- language frontier maps for zero-shot semantic navigation

    Naoki Harrison Yokoyama, Sehoon Ha, Dhruv Batra, Ji- uguang Wang, and Bernadette Bucher. Vlfm: Vision- language frontier maps for zero-shot semantic navigation. In 2nd Workshop on Language and Robot Learning: Language as Grounding, 2023. 1, 2, 3, 4

    2023

  13. [21]

    L3mvn: Leveraging large language models for visual target naviga- tion

    Bangguo Yu, Hamidreza Kasaei, and Ming Cao. L3mvn: Leveraging large language models for visual target naviga- tion. In2023 IEEE/RSJ International Conference on Intel- ligent Robots and Systems (IROS), pages 3554–3560. IEEE,

  14. [22]

    Multi- floor zero-shot object navigation policy.arXiv preprint arXiv:2409.10906, 2024

    Lingfeng Zhang, Hao Wang, Erjia Xiao, Xinyao Zhang, Qiang Zhang, Zixuan Jiang, and Renjing Xu. Multi- floor zero-shot object navigation policy.arXiv preprint arXiv:2409.10906, 2024. 4

    2024 arXiv

  15. [23]

    Trihelper: Zero- shot object navigation with dynamic assistance

    Lingfeng Zhang, Qiang Zhang, Hao Wang, Erjia Xiao, Zix- uan Jiang, Honglei Chen, and Renjing Xu. Trihelper: Zero- shot object navigation with dynamic assistance. In2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 10035–10042. IEEE, 2024. 4

    2024

  16. [24]

    Esc: Ex- ploration with soft commonsense constraints for zero-shot object navigation

    Kaiwen Zhou, Kaizhi Zheng, Connor Pryor, Yilin Shen, Hongxia Jin, Lise Getoor, and Xin Eric Wang. Esc: Ex- ploration with soft commonsense constraints for zero-shot object navigation. InInternational Conference on Machine Learning, pages 42829–42842. PMLR, 2023. 1, 2, 3, 4 Sem...

    2023

  17. [25]

    Some example images with their corresponding GPT-4o responses

    Probability outputs from GPT-4o Figure 4. Some example images with their corresponding GPT-4o responses

  18. [26]

    This is an example where we see that SemNav quickly gets out of the room to find a couch than VLFM

    Comparison between SemNav and VLFM Figure 5. This is an example where we see that SemNav quickly gets out of the room to find a couch than VLFM. Top row: At timestep t=4, both are looking at a doorway which can lead to finding a couch. At that position SemNav had a higher prob...

  19. [27]

    Comparison among VLFM, SemNav- Greedy and SemNav Figure 6. Row 1: At timestept= 20, SemNav (with LSP) selects a frontier with a lower score (0.29 in map) to minimize expected travel cost, choosing closer frontiers that guide it toward the target. In contrast, SemNav-Greedy sel...

  20. [28]

    Here, we show a problem with the simulator where the agent successfully reached the target object, but it was still considered failed episodes

    Simulator issue Figure 7. Here, we show a problem with the simulator where the agent successfully reached the target object, but it was still considered failed episodes

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.