Pith. sign in

REVIEW 3 major objections 4 minor 3 cited by

This paper claims that in modular RL-based object-goal navigation, perception quality and test-time strategies are the decisive drivers of performance, while policy improvements with current methods yield only marginal gains.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-04 12:47 UTC pith:Q67JM64Y

load-bearing objection Perception and test-time ablations are the real contribution; the headline 'policy is marginal' is confounded by a test-time variable in Table 2, but the paper still deserves a serious referee. the 3 major comments →

arxiv 2510.01830 v2 pith:Q67JM64Y submitted 2025-10-02 cs.RO

What Matters in RL-Based Methods for Object-Goal Navigation? An Empirical Study and A Unified Framework

Hongze Wang , Boyang Sun , Jiaxu Xing , Fan Yang , Marco Hutter , Dhruv Shah , Davide Scaramuzza , Marc Pollefeys This is my paper
classification cs.RO
keywords object-goal navigationmodular reinforcement learningempirical ablation studyperception modulepolicy designtest-time enhancementGibson benchmarksemantic map
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper tries to establish which components of a modular reinforcement-learning object-goal navigation system actually drive performance. It decomposes the pipeline into perception, policy, and test-time enhancement, and runs controlled single-variable ablations over design choices in each. The authors find that improving the object detector, map size, and map construction, and adding plug-and-play heuristics at evaluation time, produce larger gains than changing the policy's observation space, action space, network architecture, or reward. Guided by these findings, they assemble a system that reaches 85.3% success and 47.5% SPL (success weighted by path length) on the Gibson benchmark, a standard indoor navigation test, surpassing previous published results, and report a human expert baseline of 98% success under identical conditions. A sympathetic reader would care because the result implies the field's current emphasis on policy-network design deserves rebalancing toward perception and inference-time engineering.

Core claim

The paper establishes that for modular RL-based ObjectNav, the ranking of importance is perception, then test-time enhancement, then policy. Controlled ablations over 1,000 episodes show a fine-tuned detector lifts success by 6.8 points on a heuristic goal policy; map augmentation and larger maps help under tighter time budgets; the three test-time heuristics (untrapping helper, dynamic goal selection, remapping mask) each improve results without retraining. Swapping CNN for Transformer, compressing the observation, or changing reward schemes moves success by less than a point. The assembled best-configuration policies reach 85.3% success and 47.5% SPL (success weighted by path length) on th

What carries the argument

The central mechanism is a modular decomposition and a controlled ablation protocol. Perception builds a top-down semantic map from RGB-D and pose using an object detector (Mask R-CNN, an instance segmentation model, or RedNet, an RGB-D segmentation network), with map size and map augmentation as variables. The policy consumes that map and predicts a goal via an observation space (raw or compressed RGB), an action space (continuous or discrete corner goals), a backbone (CNN or Transformer), and a reward. Test-time enhancement adds three plug-and-play heuristics: an untrapping helper (alternating turns after collisions), dynamic goal selection (switch goals only when reached or unreachable),

Load-bearing premise

The conclusion that policy improvements are marginal rests on the assumption that the fixed 2-million-frame, single-seed training protocol is a fair arena for comparing policy designs; if a stronger policy would need more training or multiple seeds to reveal its advantage, the headline ranking of what matters could invert.

What would settle it

Train the same two policy architectures (CNN vs Transformer) on identical perception and test-time modules for 10 million frames across three seeds and plot convergence. If the Transformer exceeds the CNN by more than a few success-rate points at convergence, the claim that policy improvements are marginal is falsified. A complementary check: rerun the perception ablations with a near-perfect detector; if policy choices then drive large differences, the perception-dominance result is an artifact of detector quality.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Domain-aligned object detection is the highest-leverage investment: a Gibson-finetuned detector added 6.8 success-rate points on a heuristic goal policy, and mis-detection was the largest single failure category (68 of 120 failures).
  • Plug-and-play test-time heuristics can deliver large, retraining-free gains: untrapping helper, dynamic goal selection, and remapping mask all improve success and efficiency, especially in long-horizon episodes.
  • Compressing the semantic-map observation to a 3-channel RGB image preserves success within 0.5 points while cutting the observation dimension by more than sixfold, so efficiency can be gained without a performance penalty.
  • For policy design, continuous action spaces and a distance-to-target reward term are preferable to discrete corner goals and exploration-only rewards.
  • Human-level performance (98% success) is far above current RL agents, so the remaining headroom lies mainly in perception robustness and exploration strategy rather than policy-network complexity.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The 'policy gains are marginal' conclusion is conditioned on a 2-million-frame, single-seed training protocol; a longer or multi-seed comparison could reveal larger policy-architecture differences, so the ranking should be read as true for current training budgets rather than as a universal law.
  • Because the test-time heuristics are hand-tuned to simulator failure modes (staircase overlap, narrow-space trapping), their decisive contribution may not transfer to physical robots or open-floor layouts, limiting the universality of the perception/test-time dominance claim beyond simulation.
  • The success of compressing semantic maps to RGB images suggests a broader opportunity: observation-space information density may be a cheap way to trade memory for robustness, and this idea could be tested on end-to-end navigation agents as well.
  • If detector accuracy is the primary bottleneck, a natural extension is to measure navigation success as a function of detector recall for rare object categories; the failure-case counts predict a near-linear sensitivity that could be checked directly.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper conducts a large-scale empirical study of modular RL-based Object-Goal Navigation on Gibson, decomposing systems into perception, policy, and test-time enhancement modules. Through controlled ablations it claims that perception quality and test-time strategies are decisive, while policy improvements are only marginal. Based on these findings, it proposes a unified framework, introduces dynamic evaluation metrics, reports state-of-the-art results (85.3% SR, 47.5% SPL), and provides a human expert baseline at 98% SR.

Significance. If the central claims hold, the paper provides a valuable reallocation of research effort in ObjectNav: away from policy-network design and toward perception and inference-time engineering. The study is unusually broad, covers many design choices with a consistent testbed, and includes a human baseline and planned code release. However, the headline conclusions rest on the internal validity of the policy-module ablations and on statistical reliability of comparisons; both are currently problematic.

major comments (3)
  1. [§4.2, Table 2 vs Table 3] The policy-module ablations are not controlled for test-time enhancements. Table 2 row 3 (continuous, SR 84.4, SPL 48.2) exactly matches Table 3's continuous policy with untrapping helper, dynamic goal selection, and remapping mask all enabled. Table 2 row 2 (discrete, SR 83.8, SPL 43.3) exactly matches Table 3's discrete policy with dynamic goal selection disabled (row 4). Thus the action-space comparison in §4.2.2 simultaneously changes dynamic goal selection. Adding dynamic goal selection to the discrete policy (Table 3 row 5) yields SR 85.3 / SPL 47.5, statistically indistinguishable from the continuous policy. The claimed 4.9 SPL gain is therefore not attributable to the action space, and the claim that policy advances are marginal is not cleanly measured. The ablations should be rerun with a fixed test-time enhancement configuration, or the conclusions must be re-scoped.
  2. [Appendix A.2, Tables 2–3] All results come from a single training run per configuration, with no seeds, error bars, or significance tests. Several headline comparisons are within 1% (observation space <0.5% SR, architecture 0.1% SR), which is within the typical seed noise of RL training. The fixed 2M-frame budget is justified by citing prior work but no convergence curves are shown; a Transformer policy that would overtake a CNN with more training or different reward tuning would invert the 'policy improvements are marginal' conclusion. The paper should report at least 3 seeds and learning curves, or explicitly present the conclusions as indicative rather than decisive.
  3. [§4.2.4, Table 2] The reward-design comparison is reported as a 0.8% SR / 2.5% SPL improvement, but the actual rows compared are not clearly identified. Row 5 (Transformer, Type 1) vs row 2 (Transformer, Type 2) gives 1.4% SR / 2.5% SPL. Row 5 vs row 4 (CNN, Type 2) changes both architecture and reward. Since Section 4.2 states that all other factors are held constant, the reward comparison must be explicitly defined and must come from rows that differ only in reward. The current table does not support the stated numbers unambiguously.
minor comments (4)
  1. [§A.3] The dynamic timestep metric depends on the free scaling factor alpha, which is set to 5 'to better distinguish performance differences.' A sensitivity analysis over alpha is needed to show that conclusions are not an artifact of this choice.
  2. [§4.4 / Appendix A.7] The human baseline is described as 'under identical conditions,' but humans practice on training scenes and the test subset is chosen so the policy performs similarly. These differences should be acknowledged in the main text, not only in the appendix.
  3. [§4.1.2 / Figure 8] For the map-size analysis, the corner-goal policy is said to benefit from larger maps because goals are placed at corners; the generalization of this recommendation to other goal-selection schemes should be stated more cautiously, since the frontier policy shows the opposite trend.
  4. [Throughout] Several small claims are made without direct evidence, e.g., 'the field still lacks a unifying analysis' and 'the first large-scale empirical study.' These should be softened or supported with a systematic comparison to prior ablation studies.

Circularity Check

0 steps flagged

No significant circularity; the study is an external empirical benchmark comparison with no self-citations and no result that reduces to its own inputs.

full rationale

This paper is an empirical study rather than a derivation chain: it reports controlled experiments and compares against external published baselines on the Gibson benchmark. There are no author self-citations carrying the argument (the reference list contains no work by the present authors), no fitted parameters are renamed as predictions, and no uniqueness or ansatz is imported from the authors' prior work. The central conclusions about perception, policy, and test-time enhancement are supported by table ablations and comparisons with prior methods, not by equations that are reused as results. Two caveats are worth noting but they are correctness/validity concerns rather than circularity: (1) Appendix A.3 states alpha=5 was chosen 'to better distinguish the performance differences among various methods,' which means the dynamic-metric magnitudes are partly a measurement-design choice, but this does not make any component's contribution equivalent to its input. (2) The continuous-action row in Table 2 matches the all-test-time-enhancements configuration in Table 3, while the discrete-action row has dynamic goal selection disabled, so the action-space comparison is confounded with a test-time enhancement; this is an experimental-control flaw, not a definitional or self-referential reduction. No result in the paper is shown to be equivalent to its own inputs by construction, so the circularity score is 0.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The central claims rest on few fitted quantities: the hand-calibrated dynamic-timestep factor alpha=5 and several unreported policy and trick thresholds. The 2M-frame single-seed training protocol is the most consequential unstated premise. No new physical entities or forces are introduced; the three test-time strategies are heuristics borrowed from prior work (Stubborn, FSE), and the 'unified framework' is a software scaffold.

free parameters (4)
  • Dynamic timestep scaling factor alpha = 5
    Chosen in Appendix A.3 'to better distinguish the performance differences among various methods'; larger alpha loosens the step budget, smaller tightens it. The value is hand-calibrated to amplify inter-method separation in the proposed D-SR/D-SPL metrics.
  • Reward weights alpha_1, alpha_2, alpha_3, alpha_4 = not stated in text
    Defined symbolically in Appendix A.5 for exploration, distance-to-target, success bonus, and step penalty; numeric values are not reported, so re-implementation requires guessing.
  • Untrapping helper thresholds tau_coll, tau_block = not stated in text
    Algorithm 1 in Appendix A.6 defines collision/block thresholds symbolically; values are absent, and they gate whether the helper triggers during evaluation.
  • Dynamic goal selection hyperparameters f_update, tau_unreachable, tau_reached = not stated in text
    Algorithm 2 in Appendix A.6 relies on goal update frequency and unreachable/reached thresholds; no values are given, and the strategy is shown to significantly affect performance (Table 3).
axioms (4)
  • domain assumption 2 million training frames is a fair regime for all compared policies (convergence)
    Appendix A.2 states all experiments were trained for 2M frames, citing prior works trained 0.1M-10M by saying 'around 1 million frames typically yielding good performance'; no convergence curves per policy architecture are shown. If the Transformer policy is undertrained, the 'policy gains are marginal' conclusion (Table 2, Section 4.2.3) is an artifact.
  • domain assumption Published numbers for prior methods (Table 4) come from protocols matched to the authors' evaluation
    Section 4.4 compares SemExp/PONI/FSE/NaviFormer/T-Diff/HOZ++ using numbers from their papers; the paper does not re-run prior methods in its own Habitat configuration, so episode sampling, sensor noise, action frequencies, and success radius are assumed equivalent.
  • domain assumption The within-1-meter success criterion inherited from Habitat is an adequate measure of ObjectNav success
    Section 3 defines success as being within 1m of the target at episode end; the failure analysis (Appendix A.1) itself notes 21 episodes humans judged successful but the simulator marked failed because of this strict criterion, so the metric is known to be strict and the authors do not correct for it.
  • domain assumption Per-module marginal conclusions are not inverted by interaction effects
    Section 3 decomposes the system into three modules and studies each while fixing others; the paper itself reports at least one interaction (map size effects differ for corner vs frontier policies, Section 4.1.2), so the per-module conclusions are settings-dependent rather than universal laws.

pith-pipeline@v1.3.0-alltime-deepseek · 16482 in / 21558 out tokens · 155366 ms · 2026-08-04T12:47:48.053261+00:00 · methodology

0 comments
read the original abstract

Object-Goal Navigation (ObjectNav) is a key capability for deploying mobile robots in everyday environments such as homes, schools, and workplaces. In this task, an agent must locate an instance of a target object category in previously unseen environments using only onboard perception, requiring the integration of semantic understanding, spatial reasoning, and long-horizon planning. Reinforcement learning (RL) has become a dominant paradigm for ObjectNav, yet modern systems involve numerous design choices across perception modules, policy architectures, and inference-time strategies. The relative impact of these components, however, remains poorly understood. In this work, we present a large-scale empirical study of modular RL-based ObjectNav systems. We decompose the navigation pipeline into three key components: perception, policy, and test-time enhancement, and conduct extensive controlled experiments to analyze their individual contributions. Our results suggest that improvements in perception quality and test-time strategies often yield larger performance gains than policy improvements alone, highlighting the importance of understanding how different components interact within modular navigation systems. Motivated by these findings, we introduce a unified framework for systematically studying modular ObjectNav systems. Guided by our analysis, we build an enhanced system that achieves state-of-the-art performance on the Gibson benchmark, improving SPL by 6.6% and success rate by 2.7% over prior methods. We also introduce a human expert baseline, achieving 98% success, highlighting the significant gap between current RL agents and human-level navigation. Finally, we provide practical insights and design recommendations for each module to help guide future research. Project page: https://honwang0054.github.io/What-matters-in-RL-ObjNav-web/.

Figures

Figures reproduced from arXiv: 2510.01830 by Boyang Sun, Davide Scaramuzza, Dhruv Shah, Fan Yang, Hongze Wang, Jiaxu Xing, Marco Hutter, Marc Pollefeys.

Figure 1
Figure 1. Figure 1: Overview of our work. Our framework encompasses: (1) an empirical study analyzing [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Unified Framework of our experimental set￾ting. A. Perception: RGB-D and pose are fused into a top-down semantic map. B. Policy: The map and aux￾iliary inputs (e.g., category, orientation) guide action prediction. C. Test-Time Enhancement: Plug-and-play strategies applied at evaluation to boost performance without retraining. System Overview [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Analysis of Test Navigation Scenarios. Sankey plots illustrate the distribution of success and failure cases over 1,000 test episodes across five indoor scenes. We analyze the test results over 1,000 episodes across five different test scenes, manually examining each case to identify failure points, with 200 episodes evaluated for each policy. From [PITH_FULL_IMAGE:figures/full_fig_p014_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Perception Module Overview. RGB images are processed by a pretrained object de￾tector for semantic labels, which are projected with depth-based point clouds to form a voxel map. Summing across height levels yields a multi-layer top-down semantic map, where K is the channel number and M the map size. Object Detector The first step in the perception module is to extract semantic categories from the input RGB… view at source ↗
Figure 5
Figure 5. Figure 5: Action Space. The current goal position is represented by a blue filled dot or line. For both continuous and discrete action spaces, red unfilled dots indicate the possible next goal positions. As shown in the map, in the continuous action space, the next goal can be located anywhere on the map. In contrast, for the discrete action space, the next goal is selected only from a predefined list of candidate p… view at source ↗
Figure 6
Figure 6. Figure 6: ViT based Policy Network. Our RL policy is adapted from a pretrained Vision Trans￾former (ViT). The compressed top-down semantic map is divided into 16 patches and passed through the transformer encoder, where the [CLS] token output is used to predict both action and value. Reward Design For the ObjectGoal navigation task, we consider four types of rewards: exploration, distance-based, success bonus, and s… view at source ↗
Figure 7
Figure 7. Figure 7: Comparison of Different Object Detector Models in Habitat the corners of the map, a larger map effectively sets more distant goals. This encourages the agent to explore a wider area, which can lead to improved navigation performance. This trend does not hold for the Frontier-Based Policy, where actions target frontiers of the explored space and are less sensitive to map size. In some cases, higher-resoluti… view at source ↗
Figure 8
Figure 8. Figure 8: Comparison of map sizes in Habitat. Top row shows maps of 240 × 240. Bottom row shows 480 × 480. 19 [PITH_FULL_IMAGE:figures/full_fig_p019_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Comparison of map augmentation in Habitat. Top row: augmented maps. Bottom row: non-augmented. Blue dashed boxes highlight key differences. 20 [PITH_FULL_IMAGE:figures/full_fig_p020_9.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. LocalNav: Distilling Frontier VLMs and Embodied RL for On-Device Object Goal Navigation

    cs.RO 2026-06 unverdicted novelty 6.0

    Distillation from frontier VLMs plus E-RLVR regularization produces a 4B local model that achieves 34.5% SR on OVON while cutting inference latency by 82.8%.

  2. OpenFrontier: General Navigation with Visual-Language Grounded Frontiers

    cs.RO 2026-03 unverdicted novelty 5.0

    OpenFrontier formulates robot navigation as sparse subgoal reaching via visual-language-grounded frontiers, achieving zero-shot performance without fine-tuning or dense semantic maps.

  3. OpenFrontier: General Navigation with Visual-Language Grounded Frontiers

    cs.RO 2026-03 unverdicted novelty 5.0

    OpenFrontier treats navigation as sparse visual-frontier subgoal selection guided by vision-language priors, claiming strong zero-shot and real-robot performance without task-specific training.

Reference graph

Works this paper leans on

43 extracted references · 1 canonical work pages · cited by 2 Pith papers

  1. [1]

    Peter Anderson, Angel Chang, Devendra Singh Chaplot, Alexey Dosovitskiy, Saurabh Gupta, Vladlen Koltun, Jana Kosecka, Jitendra Malik, Roozbeh Mottaghi, Manolis Savva, and Amir R. Zamir. On evaluation of embodied navigation agents, 2018. URL https://arxiv.org/abs/1807.06757

    Pith/arXiv arXiv 2018

  2. [2]

    Matterport3d: Learning from rgb-d data in indoor environments

    Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments. International Conference on 3D Vision (3DV), 2017

    2017

  3. [3]

    Object goal navigation using goal-oriented semantic exploration

    Devendra Singh Chaplot, Dhiraj Gandhi, Abhinav Gupta, and Ruslan Salakhutdinov. Object goal navigation using goal-oriented semantic exploration. In In Neural Information Processing Systems, 2020 a

    2020

  4. [4]

    Learning to explore using active neural slam

    Devendra Singh Chaplot, Dhiraj Gandhi, Saurabh Gupta, Abhinav Gupta, and Ruslan Salakhutdinov. Learning to explore using active neural slam. In International Conference on Learning Representations (ICLR), 2020 b

    2020

  5. [5]

    Masked-attention mask transformer for universal image segmentation

    Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 1290--1299, 2022

    2022

  6. [6]

    Spoc: Imitating shortest paths in simulation enables effective navigation and manipulation in the real world, 2024

    Kiana Ehsani, Tanmay Gupta, Rose Hendrix, Jordi Salvador, Luca Weihs, Kuo-Hao Zeng, Kunal Pratap Singh, Yejin Kim, Winson Han, Alvaro Herrasti, Ranjay Krishna, Dustin Schwenk, Eli VanderBilt, and Aniruddha Kembhavi. Spoc: Imitating shortest paths in simulation enables effective navigation and manipulation in the real world, 2024. URL https://arxiv.org/abs...

    Pith/arXiv arXiv 2024

  7. [7]

    Navigating to objects in the real world

    Theophile Gervet, Soumith Chintala, Dhruv Batra, Jitendra Malik, and Devendra Singh Chaplot. Navigating to objects in the real world. Science Robotics, 8 0 (79): 0 eadf6991, 2023

    2023

  8. [8]

    Object goal visual navigation using semantic spatial relationships

    Jingwen Guo, Zhisheng Lu, Ti Wang, Weibo Huang, and Hong Liu. Object goal visual navigation using semantic spatial relationships. In CAAI international conference on artificial intelligence, pp.\ 77--88. Springer, 2021

    2021

  9. [9]

    Zhang, Shaoqing Ren, and Jian Sun

    Kaiming He, X. Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp.\ 770--778, 2015. URL https://api.semanticscholar.org/CorpusID:206594692

    2016

  10. [10]

    Mask r-cnn

    Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask r-cnn. In 2017 IEEE International Conference on Computer Vision (ICCV), pp.\ 2980--2988, 2017. doi:10.1109/ICCV.2017.322

    doi:10.1109/iccv.2017.322 2017

  11. [11]

    Oneformer: One transformer to rule universal image segmentation

    Jitesh Jain, Jiachen Li, Mang Tik Chiu, Ali Hassani, Nikita Orlov, and Humphrey Shi. Oneformer: One transformer to rule universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 2989--2998, 2023

    2023

  12. [12]

    Rednet: Residual encoder-decoder network for indoor rgb-d semantic segmentation, 2018

    Jindong Jiang, Lunan Zheng, Fei Luo, and Zhijun Zhang. Rednet: Residual encoder-decoder network for indoor rgb-d semantic segmentation, 2018. URL https://arxiv.org/abs/1806.01054

    Pith/arXiv arXiv 2018

  13. [13]

    Simple but effective: Clip embeddings for embodied ai

    Apoorv Khandelwal, Luca Weihs, Roozbeh Mottaghi, and Aniruddha Kembhavi. Simple but effective: Clip embeddings for embodied ai. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2022

    2022

  14. [14]

    Berg, Wan-Yen Lo, Piotr Doll \'a r, and Ross Girshick

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Doll \'a r, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023

    Pith/arXiv arXiv 2023

  15. [15]

    Stubborn: A strong baseline for indoor object navigation, 2022

    Haokuan Luo, Albert Yue, Zhang-Wei Hong, and Pulkit Agrawal. Stubborn: A strong baseline for indoor object navigation, 2022. URL https://arxiv.org/abs/2203.07359

    Pith/arXiv arXiv 2022

  16. [16]

    Turner, Oleksandr Maksymets, Zsolt Kira, Mrinal Kalakrishnan, Jitendra Malik, Devendra Singh Chaplot, Unnat Jain, Dhruv Batra, Akshara Rai, and Roozbeh Mottaghi

    Xavier Puig, Eric Undersander, Andrew Szot, Mikael Dallaire Cote, Tsung-Yen Yang, Ruslan Partsey, Ruta Desai, Alexander William Clegg, Michal Hlavac, So Yeon Min, Vladimír Vondruš, Theophile Gervet, Vincent-Pierre Berges, John M. Turner, Oleksandr Maksymets, Zsolt Kira, Mrinal Kalakrishnan, Jitendra Malik, Devendra Singh Chaplot, Unnat Jain, Dhruv Batra, ...

    Pith/arXiv arXiv 2023

  17. [17]

    Christensen

    Yiding Qiu, Anwesan Pal, and Henrik I. Christensen. Learning hierarchical relationships for object-goal navigation, 2020. URL https://arxiv.org/abs/2003.06749

    Pith/arXiv arXiv 2020

  18. [18]

    Poni: Potential functions for objectgoal navigation with interaction-free learning, 2022

    Santhosh Kumar Ramakrishnan, Devendra Singh Chaplot, Ziad Al-Halah, Jitendra Malik, and Kristen Grauman. Poni: Potential functions for objectgoal navigation with interaction-free learning, 2022. URL https://arxiv.org/abs/2201.10029

    Pith/arXiv arXiv 2022

  19. [19]

    Habitat-web: Learning embodied object-search strategies from human demonstrations at scale, 2022

    Ram Ramrakhya, Eric Undersander, Dhruv Batra, and Abhishek Das. Habitat-web: Learning embodied object-search strategies from human demonstrations at scale, 2022. URL https://arxiv.org/abs/2204.03514

    Pith/arXiv arXiv 2022

  20. [20]

    Pirlnav: Pretraining with imitation and rl finetuning for objectnav, 2023

    Ram Ramrakhya, Dhruv Batra, Erik Wijmans, and Abhishek Das. Pirlnav: Pretraining with imitation and rl finetuning for objectnav, 2023. URL https://arxiv.org/abs/2301.07302

    Pith/arXiv arXiv 2023

  21. [21]

    Sam 2: Segment anything in images and videos

    Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman R \"a dle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Doll \'a r, and Christoph Feichtenhofer. Sam 2: Segment anything in images and videos. arXiv preprint arXiv...

    Pith/arXiv arXiv 2024

  22. [23]

    Visual transformers: Token-based image representation and processing for computer vision, 2020

    Bichen Wu, Chenfeng Xu, Xiaoliang Dai, Alvin Wan, Peizhao Zhang, Zhicheng Yan, Masayoshi Tomizuka, Joseph Gonzalez, Kurt Keutzer, and Peter Vajda. Visual transformers: Token-based image representation and processing for computer vision, 2020

    2020

  23. [24]

    Detectron2

    Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detectron2. https://github.com/facebookresearch/detectron2, 2019

    2019

  24. [25]

    Zamir, Zhi-Yang He, Alexander Sax, Jitendra Malik, and Silvio Savarese

    Fei Xia, Amir R. Zamir, Zhi-Yang He, Alexander Sax, Jitendra Malik, and Silvio Savarese. Gibson env: real-world perception for embodied agents. In Computer Vision and Pattern Recognition (CVPR), 2018 IEEE Conference on. IEEE, 2018

    2018

  25. [26]

    Naviformer: A spatio-temporal context-aware transformer for object navigation

    Wei Xie, Haobo Jiang, Yun Zhu, Jianjun Qian, and Jin Xie. Naviformer: A spatio-temporal context-aware transformer for object navigation. Proceedings of the AAAI Conference on Artificial Intelligence, 39 0 (14): 0 14708--14716, Apr. 2025. doi:10.1609/aaai.v39i14.33612. URL https://ojs.aaai.org/index.php/AAAI/article/view/33612

    work page doi:10.1609/aaai.v39i14.33612 2025

  26. [27]

    Offline visual representation learning for embodied navigation, 2022

    Karmesh Yadav, Ram Ramrakhya, Arjun Majumdar, Vincent-Pierre Berges, Sachit Kuhar, Dhruv Batra, Alexei Baevski, and Oleksandr Maksymets. Offline visual representation learning for embodied navigation, 2022. URL https://arxiv.org/abs/2204.13226

    Pith/arXiv arXiv 2022

  27. [28]

    A frontier-based approach for autonomous exploration

    Brian Yamauchi. A frontier-based approach for autonomous exploration. Proceedings 1997 IEEE International Symposium on Computational Intelligence in Robotics and Automation CIRA'97. 'Towards New Computational Principles for Robotics and Automation', pp.\ 146--151, 1997. URL https://api.semanticscholar.org/CorpusID:206561621

    1997

  28. [29]

    Visual semantic navigation using scene priors, 2018

    Wei Yang, Xiaolong Wang, Ali Farhadi, Abhinav Gupta, and Roozbeh Mottaghi. Visual semantic navigation using scene priors, 2018. URL https://arxiv.org/abs/1810.06543

    Pith/arXiv arXiv 2018

  29. [30]

    Auxiliary tasks and exploration enable objectnav, 2021

    Joel Ye, Dhruv Batra, Abhishek Das, and Erik Wijmans. Auxiliary tasks and exploration enable objectnav, 2021. URL https://arxiv.org/abs/2104.04112

    Pith/arXiv arXiv 2021

  30. [31]

    Frontier semantic exploration for visual target navigation

    Bangguo Yu, Hamidreza Kasaei, and Ming Cao. Frontier semantic exploration for visual target navigation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pp.\ 4099–4105. IEEE, May 2023 a . doi:10.1109/icra48891.2023.10161059. URL http://dx.doi.org/10.1109/ICRA48891.2023.10161059

    arXiv 2023

  31. [32]

    L3mvn: Leveraging large language models for visual target navigation

    Bangguo Yu, Hamidreza Kasaei, and Ming Cao. L3mvn: Leveraging large language models for visual target navigation. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp.\ 3554–3560. IEEE, October 2023 b . doi:10.1109/iros55552.2023.10342512. URL http://dx.doi.org/10.1109/IROS55552.2023.10342512

    arXiv 2023

  32. [33]

    Trajectory diffusion for objectgoal navigation

    Xinyao Yu, Sixian Zhang, Xinhang Song, Xiaorong Qin, and Shuqiang Jiang. Trajectory diffusion for objectgoal navigation. In A. Globerson, L. Mackey, D. Belgrave, A. Fan, U. Paquet, J. Tomczak, and C. Zhang (eds.), Advances in Neural Information Processing Systems, volume 37, pp.\ 110388--110411. Curran Associates, Inc., 2024. URL https://proceedings.neuri...

    2024

  33. [34]

    Poliformer: Scaling on-policy rl with transformers results in masterful navigators, 2024

    Kuo-Hao Zeng, Zichen Zhang, Kiana Ehsani, Rose Hendrix, Jordi Salvador, Alvaro Herrasti, Ross Girshick, Aniruddha Kembhavi, and Luca Weihs. Poliformer: Scaling on-policy rl with transformers results in masterful navigators, 2024. URL https://arxiv.org/abs/2406.20083

    Pith/arXiv arXiv 2024

  34. [35]

    PEANUT : Predicting and navigating to unseen targets

    Albert J Zhai and Shenlong Wang. PEANUT : Predicting and navigating to unseen targets. In ICCV, 2023

    2023

  35. [36]

    3d-aware object goal navigation via simultaneous exploration and identification, 2023

    Jiazhao Zhang, Liu Dai, Fanpeng Meng, Qingnan Fan, Xuelin Chen, Kai Xu, and He Wang. 3d-aware object goal navigation via simultaneous exploration and identification, 2023. URL https://arxiv.org/abs/2212.00338

    Pith/arXiv arXiv 2023

  36. [37]

    Hierarchical object-to-zone graph for object navigation, 2021

    Sixian Zhang, Xinhang Song, Yubing Bai, Weijie Li, Yakui Chu, and Shuqiang Jiang. Hierarchical object-to-zone graph for object navigation, 2021. URL https://arxiv.org/abs/2109.02066

    Pith/arXiv arXiv 2021

  37. [38]

    Imagine before go: Self-supervised generative map for object goal navigation

    Sixian Zhang, Xinyao Yu, Xinhang Song, Xiaohan Wang, and Shuqiang Jiang. Imagine before go: Self-supervised generative map for object goal navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp.\ 16414--16425, June 2024

    2024

  38. [39]

    Hoz++: Versatile hierarchical object-to-zone graph for object navigation

    Sixian Zhang, Xinhang Song, Xinyao Yu, Yubing Bai, Xinlong Guo, Weijie Li, and Shuqiang Jiang. Hoz++: Versatile hierarchical object-to-zone graph for object navigation. IEEE Transactions on Pattern Analysis and Machine Intelligence, pp.\ 1--18, 2025. doi:10.1109/TPAMI.2025.3552987

    arXiv 2025

  39. [40]

    Visual navigation via reinforcement learning and relational reasoning

    Kang Zhou, Chi Guo, and Huyin Zhang. Visual navigation via reinforcement learning and relational reasoning. In 2021 IEEE SmartWorld, Ubiquitous Intelligence & Computing, Advanced & Trusted Computing, Scalable Computing & Communications, Internet of People and Smart City Innovation (SmartWorld/SCALCOM/UIC/ATC/IOP/SCI), pp.\ 131--138, 2021. doi:10.1109/SWC5...

    arXiv 2021

  40. [41]

    @esa (Ref

    \@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should not add it explicitly Type <Return> for now, but then later remove the command n...

  41. [42]

    \@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@first@sw \@firstoftwo \@ifundefined NAT@b*@#2 \@firstoftwo @num @NAT@ctr \@secondoft...

  42. [43]

    O. Ronneberger and P.Fischer and T. Brox

    @open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifxundefined @sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifxundefined @heading @heading NAT@ctr thebibliography [1] @ \@biblabel @NAT@ctr \@bibsetup #1 @NAT@ctr @ @openbib .11em \@plus.33em \@minus.07em 4000 4000 `\.\@m @bibit...

    Pith/arXiv arXiv 2007

  43. [44]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...