Pith. sign in

REVIEW 4 major objections 5 minor 33 references

This paper claims that a deep reinforcement learning agent can explore unknown indoor scenes more effectively—finding more objects and gravitating toward semantically rich regions—when its reward is a weighted sum of geometric novelty, obje

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 ↗

A curriculum-trained DRL agent with a VLM query action and layered rewards is claimed to improve semantic exploration and object discovery in AI2-THOR.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection Plausible VLM-as-action architecture, but the experiments never compare against any baseline, so the strong claims in the abstract are unsupported. the 4 major comments →

arxiv 2509.09356 v1 pith:JNCR34KV submitted 2025-09-11 cs.AI cs.RO

Curriculum-Based Multi-Tier Semantic Exploration via Deep Reinforcement Learning

Abdel Hakim Drid , Vincenzo Suriani , Daniele Nardi , Abderrezzak Debilou This is my paper
classification cs.AI cs.RO
keywords Deep Reinforcement LearningAutonomous ExplorationSemantic ExplorationVision-Language ModelCurriculum LearningLayered RewardAI2-THORObject Discovery
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 reading

The paper claims that an exploration agent in a simulated indoor environment performs better when its reward function is layered into three signals: new geometric keypoints, newly detected object classes, and a vision-language model's judgment of how useful a scene is. It argues that these signals cannot be learned stably all at once, so it introduces a three-phase curriculum that adds geometric, then object, then semantic rewards. It further argues that the VLM should not be queried every timestep; instead, querying is an explicit action the agent chooses, so the agent learns when external semantic guidance is worth the cost. If the paper is right, resource-constrained robots can embed common-sense semantic reasoning into exploration policies without constant expensive VLM inference.

Core claim

The central claim is that integrating VLM common sense through an action-conditioned layered reward—rather than as a passive feature extractor or dense reward—lets a DDPG-based agent learn a policy that navigates toward semantically rich regions and strategically requests VLM guidance. Concretely, the geometric layer rewards novel keypoints in an egocentric binary map; the object layer rewards discovering new YOLO-World-detected classes; the semantic layer uses GPT-4o's scene score, discretized to -1/0/+1, and is activated only when the agent executes a dedicated 'VLM-Query' action. The curriculum trains the geometric layer first, then adds object detection, then semantic scoring. The paper

What carries the argument

The layered reward function is the central object: r_t = alpha * r_geom + beta * r_obj + delta * r_semantic. The geometric layer maintains a per-pixel binary feature map in image coordinates; the object layer tracks novel classes from YOLO-World; the semantic layer discretizes a GPT-4o scene score into -1/0/+1. The VLM-Query action makes the semantic reward action-conditional. The curriculum schedules which layers are active: Phase 1 geometric only, Phase 2 geometric+object, Phase 3 all three. DDPG learns the policy from a 128-dimensional downsampled depth state.

Load-bearing premise

The load-bearing premise is that counting keypoints in a per-frame, pixel-coordinate binary map measures progress in exploring new space; because the map is not aligned to the world, the same visible scene can be counted as new after the agent rotates or moves.

What would settle it

Observe the geometric reward in a static AI2-THOR scene: have the agent rotate in place for several timesteps without translating. If the per-pixel feature map continues to increment its count of 'new' keypoints, then the geometric layer is rewarding egocentric image novelty rather than spatial exploration, undermining the Phase 1 curriculum. A second check: compare the binary feature map against a camera-pose-registered world map; if the per-pixel map registers many repeated keypoints as new, the proxy is miscalibrated.

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

If this is right

  • Agents trained with the full three-layer curriculum discover more objects and accumulate higher detection confidence per episode than agents trained only on geometric exploration, while traveling shorter paths.
  • Explicitly modeling VLM queries as an action lets the agent control inference cost, so the number of VLM calls becomes a learned budget rather than a fixed per-step expense.
  • Curriculum staging makes it possible to add semantic supervision without destabilizing the already-learned navigation and object-seeking behaviors.
  • In more complex scenes the learned policy increases its calls to the external detector, suggesting that guidance-seeking becomes more valuable as task difficulty rises.
  • The ablation suggests that adding an accuracy-aware reward term does not substantially help, so the simpler discovery-based object reward is sufficient.

Where Pith is reading between the lines

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

  • Editorial inference: because the geometric feature map is stored in image coordinates with no camera pose or world alignment, rotating in place may register previously seen keypoints as new. A direct test—rotating in a static scene and checking whether the geometric reward accumulates—would tell whether Phase 1 rewards true novelty or egocentric novelty. The paper does not address this.
  • Editorial inference: the metrics TDO and TCS do not distinguish discovering a new object instance from re-detecting a previously seen object; if the object memory is class-based rather than instance-based, the large TDO numbers may overstate the diversity of discovered objects.
  • Editorial inference: the action-conditioned query mechanism is a generic pattern: any expensive advisory model could be wrapped as a discrete action in an RL policy, letting the agent learn a cost-benefit tradeoff. The paper's AI2-THOR results are a proof-of-concept for that pattern.
  • Editorial note: in the Related Works section the phrase 'from exploration [CIT] to object grasping [31]' contains an unresolved placeholder citation; the claim that CLIP-based exploration has been demonstrated is not supported as written.
Share X Bluesky LinkedIn Reddit HN

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

4 major / 5 minor

Summary. The paper proposes a DRL architecture for semantic exploration in AI2-THOR. The agent receives RGB-D input, maintains a pixel-coordinate binary feature map for geometric novelty, uses YOLO-World to reward new object classes, and queries GPT-4o through a dedicated discrete action to obtain a discretized semantic reward. Training proceeds in three curriculum phases: geometric, object-aware, and semantic. Experiments report maximum path length, total detected objects, and total detection confidence across phases, plus scene-specific and ablation results. The central claim is that the layered reward and curriculum produce significantly enhanced object discovery and semantically guided exploration.

Significance. If the empirical claims held, the contribution would be meaningful: modeling VLM queries as a priced action and combining layered rewards with a curriculum is a sensible way to inject common-sense knowledge into exploration while bounding computational cost. The reward decomposition is clearly stated, and the paper explicitly identifies tunable weights. However, the current evidence does not substantiate the headline claims: there are no comparisons against reasonable baselines, no error bars or multi-seed statistics, and the geometric novelty proxy is not pose-registered. The significance is therefore conditional on substantially stronger experimental support.

major comments (4)
  1. [§5.1, Tables 1 and 2] The headline claim of 'significantly enhanced object discovery' rests on phase-to-phase comparisons of a single trained agent. There are no baseline policies (random policy, geometric-only at every phase, fixed-frequency VLM queries, or non-curriculum training), no error bars, and no significance tests. The Phase 2→3 TDO gain is 1254→1274 (~1.6%) and TCS 485.09→500.09; Max PL drops from 5.75 to 5. These changes could arise from training time, detector behavior, or scene selection. Multi-seed runs, statistical comparisons, and the baselines listed above are necessary to attribute the gains to the proposed method.
  2. [§3.1, geometric reward layer] The geometric novelty signal is a per-frame pixel-coordinate binary feature map. Without camera pose or world alignment, a rotation or small translation re-registers already-observed scene content as novel, so r_geom can reward spurious novelty. This component is used alone in Phase 1 and with α=0.25 in later phases; if the proxy is flawed, the entire curriculum is mis-specified. The paper should either register keypoints in a world/anchor frame using pose, or provide a diagnostic showing that the pixel-coordinate map does not inflate novelty under rotation and translation.
  3. [§5.3, Table 3] The ablation study omits the two claimed key components: no row removes the semantic layer, the VLM-Query action, or the curriculum. The table varies input shape and an accuracy reward, so the 'geom + obj' rows do not constitute baselines for the proposed architecture. Without an ablation that turns off the semantic layer, replaces the learned VLM-Query action with a fixed query schedule, or removes the curriculum, the contribution of each component remains untested.
  4. [§3.1 and §6] The claimed 'strategic mastery of when to prompt' is not evidenced. Section 3.1 states that consecutive VLM-Query actions are penalized, so any reasonable policy will space queries; the Discussion interprets this as a learned capability. To support the claim, the paper needs statistics on query timing, information gain per query, and a comparison against a fixed or random query schedule. Without these, the observed behavior is partly by construction.
minor comments (5)
  1. [Eq. (2), Eq. (4)] The subscript/superscript is typoed as 'r^(semantnical)_t' in Eq. (2) and repeated in the text. Please correct to 'r^(semantic)_t'.
  2. [Table 2] The column 'MC' is not defined anywhere. It appears in scene-specific results but is never explained in the text or captions.
  3. [Section 2] There is an unresolved placeholder '[CIT]' in the related work paragraph: 'from exploration [CIT] to object grasping'. This citation must be filled.
  4. [References] Reference formatting is inconsistent (e.g., 'PmLR', 'ICCV' vs 'International conference on machine learning'). Please standardize the bibliography.
  5. [§4 and §5] Reproducibility is hampered by missing hyperparameters: N_max_objects, the discretization thresholds in Eq. (4), the exact VLM prompt, the number of training episodes, and the network architecture are not specified. The reward weights α, β, δ are only given in Table 2 for selected runs.

Circularity Check

0 steps flagged

No significant circularity: the central claim is an empirical performance claim, and no prediction reduces by construction to a fitted input or self-citation.

full rationale

The paper does not contain a derivation that reduces to its inputs. The layered reward (Eq. 2) is a designed objective, not a fitted predictor; the VLM-Query action is penalized for consecutive use (Sec. 3.1: "To discourage frequent 'VLM-Query' actions, we penalize consecutive uses."), so the agent's tendency to avoid redundant queries is a consequence of reward shaping, but the paper never equates 'strategic mastery' with a metric computed from that same penalty, and no equation forces the reported TDO/TCS numbers. The curriculum phases are ablations of the reward design, not circular predictions. No parameter is fit to a subset of the evaluation data and then reported as a prediction. The only author-overlapping citation is [3] (Suriani/Nardi) in a general related-work sentence ("These models allow agents to understand their surroundings with improved human-like comprehension [18], [3]."); it is not used to justify the architecture or to forbid alternatives, so it is not load-bearing. There is an incomplete citation placeholder in Sec. 2 ("from exploration [CIT] to object grasping [31]") and a lack of baselines/error bars, but these are evidence-quality issues, not circularity. Verdict: no significant circularity.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

The central claim rests on hand-chosen reward weights and thresholds, plus strong assumptions about the semantic quality of GPT-4o scores and the geometric validity of pixel-space feature counting. No new physical entities are introduced.

free parameters (4)
  • Reward weights alpha, beta, delta = (1.0, 0.0, 0.0); (0.25, 0.75, 0.0); (0.25, 0.75, 2.0)
    Values are chosen by hand in Table 2 with no sensitivity analysis or tuning procedure reported.
  • Semantic discretization thresholds = plus/minus 0.3
    Eq. (4) discretizes VLM scores at -0.3 and +0.3; these thresholds are not justified and no robustness to the choice is shown.
  • N_max_objects = not reported
    The cap in Eq. (3) on object detection reward is not given a value.
  • Consecutive VLM-query penalty = not reported
    The paper says consecutive VLM-Query actions are penalized, but the penalty magnitude is never specified.
axioms (5)
  • domain assumption Keypoints in a pixel-coordinate binary map measure geometric novelty
    Section 3.1 defines feature_map in image coordinates without pose alignment, so the same physical scene can be counted as new after camera motion.
  • domain assumption GPT-4o semantic scores correlate with exploration value
    The semantic reward layer uses SC(I_t) as ground truth for scene desirability, but no calibration or validation of this score is provided.
  • domain assumption YOLO-World detections are accurate in AI2-THOR
    Object detection reward and the TDO/TCS metrics rely entirely on YOLO-World detector outputs.
  • domain assumption DDPG with a 128-dimensional depth state can learn a robust exploration policy
    The decision module uses DDPG on a downsampled depth vector with discretized actions; no evidence is given that this state space is sufficient for the task.
  • domain assumption AI2-THOR is representative of real indoor environments
    The conclusion claims a path toward real-world autonomy, which presumes transfer from AI2-THOR to physical scenes.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Curriculum-Based Multi-Tier Semantic Exploration via Deep Reinforcement Learning." pith.science (2026). https://pith.science/paper/JNCR34KV

@misc{pith2026250909356,
  author       = {Pith},
  title        = {Pith review of: Curriculum-Based Multi-Tier Semantic Exploration via Deep Reinforcement Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JNCR34KV}},
  note         = {Machine review of arXiv:2509.09356}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Navigating and understanding complex and unknown environments autonomously demands more than just basic perception and movement from embodied agents. Truly effective exploration requires agents to possess higher-level cognitive abilities, the ability to reason about their surroundings, and make more informed decisions regarding exploration strategies. However, traditional RL approaches struggle to balance efficient exploration and semantic understanding due to limited cognitive capabilities embedded in the small policies for the agents, leading often to human drivers when dealing with semantic exploration. In this paper, we address this challenge by presenting a novel Deep Reinforcement Learning (DRL) architecture that is specifically designed for resource efficient semantic exploration. A key methodological contribution is the integration of a Vision-Language Model (VLM) common-sense through a layered reward function. The VLM query is modeled as a dedicated action, allowing the agent to strategically query the VLM only when deemed necessary for gaining external guidance, thereby conserving resources. This mechanism is combined with a curriculum learning strategy designed to guide learning at different levels of complexity to ensure robust and stable learning. Our experimental evaluation results convincingly demonstrate that our agent achieves significantly enhanced object discovery rates and develops a learned capability to effectively navigate towards semantically rich regions. Furthermore, it also shows a strategic mastery of when to prompt for external environmental information. By demonstrating a practical and scalable method for embedding common-sense semantic reasoning with autonomous agents, this research provides a novel approach to pursuing a fully intelligent and self-guided exploration in robotics.

Figures

Figures reproduced from arXiv: 2509.09356 by Abdel Hakim Drid, Abderrezzak Debilou, Daniele Nardi, Vincenzo Suriani.

Figure 1
Figure 1. Figure 1: Architecture of our novel DRL agent, highlighting the key contribution: a [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Architecture of our novel DRL agent, highlighting the key contribution: a [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Concise state representation designed for efficient learning. (a) RGB [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Agent Trajectories Across Four Illustrative AI2-THOR Scenes. Scene 1 [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: This figure presents bar plots for three key performance metrics across [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

33 extracted references · 14 linked inside Pith

  1. [1]

    arXiv preprint arXiv:2303.08774 (2023)

    Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F.L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al.: Gpt-4 technical report. arXiv preprint arXiv:2303.08774 (2023)

    Pith/arXiv arXiv 2023

  2. [2]

    Advances in neural information processing systems35, 23716– 23736 (2022)

    Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Men- sch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems35, 23716– 23736 (2022)

    2022

  3. [3]

    In: 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)

    Argenziano, F., Brienza, M., Suriani, V., Nardi, D., Bloisi, D.D.: Empower: em- bodied multi-role open-vocabulary planning with online grounding and execution. In: 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 12040–12047. IEEE (2024)

    2024

  4. [4]

    Advances in neural information processing systems33, 1877–1901 (2020)

    Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Nee- lakantan, A., Shyam, P., Sastry, G., Askell, A., et al.: Language models are few-shot learners. Advances in neural information processing systems33, 1877–1901 (2020)

    1901

  5. [5]

    arXiv preprint arXiv:2305.16925 (2023)

    Chen, J., Li, G., Kumar, S., Ghanem, B., Yu, F.: How to not train your dragon: Training-free embodied object goal navigation with semantic frontiers. arXiv preprint arXiv:2305.16925 (2023)

    Pith/arXiv arXiv 2023

  6. [6]

    arXiv preprint arXiv:2402.02651 (2024)

    Chen, W., Mees, O., Kumar, A., Levine, S.: Vision-language models pro- vide promptable representations for reinforcement learning. arXiv preprint arXiv:2402.02651 (2024)

    Pith/arXiv arXiv 2024

  7. [7]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Cheng, T., Song, L., Ge, Y., Liu, W., Wang, X., Shan, Y.: Yolo-world: Real-time open-vocabulary object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16901–16911 (2024)

    2024

  8. [8]

    cat- shaped mug

    Dorbala, V.S., Mullen, J.F., Manocha, D.: Can an embodied agent find your “cat- shaped mug”? llm-based zero-shot object navigation. IEEE Robotics and Automa- tion Letters9(5), 4083–4090 (2023)

    2023

  9. [9]

    arXiv preprint arXiv:2407.21783 (2024)

    Grattafiori, A., Dubey, A., Jauhri, A., Pandey, A., Kadian, A., Al-Dahle, A., Let- man, A., Mathur, A., Schelten, A., Vaughan, A., et al.: The llama 3 herd of models. arXiv preprint arXiv:2407.21783 (2024)

    Pith/arXiv arXiv 2024

  10. [10]

    Neural Computing and Applications36(21), 13205–13221 (2024)

    He, Y., Zhou, K.: Relation-wise transformer network and reinforcement learning for visual navigation. Neural Computing and Applications36(21), 13205–13221 (2024)

    2024

  11. [11]

    In: 2023 IEEE International Conference on Robotics and Automation (ICRA)

    Huang, C., Mees, O., Zeng, A., Burgard, W.: Visual language maps for robot navigation. In: 2023 IEEE International Conference on Robotics and Automation (ICRA). pp. 10608–10615. IEEE (2023)

    2023

  12. [12]

    In: International conference on machine learning

    Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q., Sung, Y.H., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision. In: International conference on machine learning. pp. 4904–4916. PMLR (2021)

    2021

  13. [13]

    arXiv preprint arXiv:1712.05474 (2017)

    Kolve, E., Mottaghi, R., Han, W., VanderBilt, E., Weihs, L., Herrasti, A., Deitke, M.,Ehsani,K.,Gordon,D.,Zhu,Y.,etal.:Ai2-thor:Aninteractive3denvironment for visual ai. arXiv preprint arXiv:1712.05474 (2017)

    Pith/arXiv arXiv 2017

  14. [14]

    In: International confer- ence on machine learning

    Li, J., Li, D., Xiong, C., Hoi, S.: Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International confer- ence on machine learning. pp. 12888–12900. PMLR (2022) 14 Drid et al

    2022

  15. [15]

    In: Proceed- ings of the IEEE/CVF conference on computer vision and pattern recognition

    Li, L.H., Zhang, P., Zhang, H., Yang, J., Li, C., Zhong, Y., Wang, L., Yuan, L., Zhang, L., Hwang, J.N., et al.: Grounded language-image pre-training. In: Proceed- ings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10965–10975 (2022)

    2022

  16. [16]

    arXiv preprint arXiv:1509.02971 (2015)

    Lillicrap, T.P., Hunt, J.J., Pritzel, A., Heess, N., Erez, T., Tassa, Y., Silver, D., Wierstra, D.: Continuous control with deep reinforcement learning. arXiv preprint arXiv:1509.02971 (2015)

    Pith/arXiv arXiv 2015

  17. [17]

    Advances in neural information processing systems36, 34892–34916 (2023)

    Liu, H., Li, C., Wu, Q., Lee, Y.J.: Visual instruction tuning. Advances in neural information processing systems36, 34892–34916 (2023)

    2023

  18. [18]

    In: 2024 IEEE 20th International Conference on Automation Science and Engineering (CASE)

    Liu, S., Zhang, J., Gao, R.X., Wang, X.V., Wang, L.: Vision-language model- driven scene understanding and robotic object manipulation. In: 2024 IEEE 20th International Conference on Automation Science and Engineering (CASE). pp. 21–26. IEEE (2024)

    2024

  19. [19]

    arXiv preprint arXiv:1907.11692 (2019)

    Liu, Y., Ott, M., Goyal, N., Du, J., Joshi, M., Chen, D., Levy, O., Lewis, M., Zettlemoyer, L., Stoyanov, V.: Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692 (2019)

    Pith/arXiv arXiv 1907

  20. [20]

    arXiv preprint arXiv:2403.09583 (2024)

    Ma, R., Luijkx, J., Ajanovic, Z., Kober, J.: Explorllm: Guiding exploration in re- inforcement learning with large language models. arXiv preprint arXiv:2403.09583 (2024)

    Pith/arXiv arXiv 2024

  21. [21]

    Advances in Neural Information Processing Systems35, 32340–32352 (2022)

    Majumdar, A., Aggarwal, G., Devnani, B., Hoffman, J., Batra, D.: Zson: Zero- shot object-goal navigation using multimodal goal embeddings. Advances in Neural Information Processing Systems35, 32340–32352 (2022)

    2022

  22. [22]

    In: International conference on machine learning

    Pathak, D., Agrawal, P., Efros, A.A., Darrell, T.: Curiosity-driven exploration by self-supervised prediction. In: International conference on machine learning. pp. 2778–2787. PMLR (2017)

    2017

  23. [23]

    arXiv preprint arXiv:2410.02511 (2024)

    Qu, Y., Wang, B., Jiang, Y., Shao, J., Mao, Y., Wang, C., Liu, C., Ji, X.: Choices are more important than efforts: Llm enables efficient multi-agent exploration. arXiv preprint arXiv:2410.02511 (2024)

    Pith/arXiv arXiv 2024

  24. [24]

    In: International conference on machine learning

    Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748–8763. PmLR (2021)

    2021

  25. [25]

    Ramakrishnan, S.K., Chaplot, D.S., Al-Halah, Z., Malik, J., Grauman, K.: Poni: Potentialfunctionsforobjectgoalnavigationwithinteraction-freelearning.In:Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion. pp. 18890–18900 (2022)

    2022

  26. [26]

    arXiv preprint arXiv:2305.09295 (2023)

    Shaheer, M., Millan-Romera, J.A., Bavle, H., Sanchez-Lopez, J.L., Civera, J., Voos, H.: Graph-based global robot simultaneous localization and mapping using archi- tectural plans. arXiv preprint arXiv:2305.09295 (2023)

    Pith/arXiv arXiv 2023

  27. [27]

    Journal of Intelligent & Robotic Systems109(3), 69 (2023)

    Singh, A., Raj, K., Roy, A.M.: Efficient deep learning-based semantic mapping approach using monocular vision for resource-limited mobile robots. Journal of Intelligent & Robotic Systems109(3), 69 (2023)

    2023

  28. [28]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Sucar, E., Liu, S., Ortiz, J., Davison, A.J.: imap: Implicit mapping and position- ing in real-time. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 6229–6238 (2021)

    2021

  29. [29]

    arXiv preprint arXiv:2403.05530 (2024)

    Team, G., Georgiev, P., Lei, V.I., Burnell, R., Bai, L., Gulati, A., Tanzer, G., Vin- cent,D.,Pan,Z.,Wang,S.,etal.:Gemini1.5:Unlockingmultimodalunderstanding across millions of tokens of context. arXiv preprint arXiv:2403.05530 (2024)

    Pith/arXiv arXiv 2024

  30. [30]

    arXiv preprint arXiv:2302.13971 (2023) Curriculum-Based Semantic Exploration via Deep Reinforcement Learning 15

    Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al.: Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971 (2023) Curriculum-Based Semantic Exploration via Deep Reinforcement Learning 15

    Pith/arXiv arXiv 2023

  31. [31]

    arXiv preprint arXiv:2311.05779 (2023)

    Tziafas, G., Xu, Y., Goel, A., Kasaei, M., Li, Z., Kasaei, H.: Language-guided robot grasping: Clip-based referring grasp synthesis in clutter. arXiv preprint arXiv:2311.05779 (2023)

    Pith/arXiv arXiv 2023

  32. [32]

    arXiv preprint arXiv:2401.02695 (2024)

    Wu, P., Mu, Y., Wu, B., Hou, Y., Ma, J., Zhang, S., Liu, C.: Voronav: Voronoi- based zero-shot object navigation with large language model. arXiv preprint arXiv:2401.02695 (2024)

    Pith/arXiv arXiv 2024

  33. [33]

    IEEE Robotics and Automation Letters (2024)

    Yue, L., Zhou, D., Xie, L., Zhang, F., Yan, Y., Yin, E.: Safe-vln: Collision avoidance for vision-and-language navigation of autonomous robots operating in continuous environments. IEEE Robotics and Automation Letters (2024)

    2024

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.