Pith. sign in

REVIEW 4 major objections 5 minor 3 cited by

Active Semantic Perception

T0 review · 4 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read This paper claims that a robot can explore an indoor environment faster and more accurately by building a hierarchical scene graph, using a large language model to imagine plausible unobserved regions, and steering toward waypoints that max

desk verdict A genuinely new idea—LLM-sampled scene-graph completions used as an information-gain objective for semantic exploration—with promising but not yet clean results; the human-in-the-loop rejection of LLM outputs and an unsupported real-world claim are the main soft spots. read the letter →

arxiv 2510.05430 v2 pith:2ETSQEY7 submitted 2025-10-06 cs.RO

Huayi Tang , Pratik Chaudhari This is my paper
classification cs.RO
keywords activeperceptionsemanticmappingscenegraphsinformationgainlargelanguagemodelsautonomousexplorationrobotnavigation3Dindoorenvironments
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

The paper sets out to show that active perception—choosing where to look next—should be driven by the semantics of a scene, not just its geometry. Its central proposal is to maintain a hierarchical scene graph (rooms, objects, walls, doors, and explicitly marked empty space) and to use a large language model to generate several plausible completions of the as-yet-unseen parts of that graph. Each imagined completion is treated as a hypothesis about the world, and the robot picks the next viewpoint that would most reduce the uncertainty over these hypotheses—an information-gain objective computed directly on the scene graph. In simulated apartments and on a real quadruped robot, the method recovers object and room semantics with higher F1 scores and lower graph-edit distances than frontier-based exploration or voxel-semantic exploration, and it can predict the location of a target room before the robot ever sees it. A sympathetic reader would care because this is a concrete mechanism for injecting common-sense priors about how rooms and objects are arranged into the exploration loop itself.

What carries the argument

The load-bearing object is the multi-layer scene graph G with four node types: object nodes (reconstructed and tracked from RGB-D input), room nodes (clustered from visual place features and assigned a room label), 'nothing' nodes (large confirmed-free cuboids of space), and structure nodes (walls, doors, windows, etc. fit by plane fitting). The reasoning module uses a two-stage LLM prompt to generate m complete scene graphs from the current graph, then renders future observations Y(G,x) by checking which nodes are visible and unoccluded from viewpoint x and perturbing node locations to represent uncertainty. The decision rule is the information-gain score I(x)=I_object(x)+λI_room(x), which

What would settle it

Construct two observation histories that produce the same scene graph but lead to different future observations (e.g., two rooms identical in layout but different in texture, where texture predicts what lies behind a door). The paper's equality predicts identical information gain for a viewpoint; directly computing mutual information from the raw observation histories would show different gains, falsifying the sufficiency assumption. Alternatively, run the full exploration pipeline in an environment deliberately chosen to violate LLM priors (e.g., a 'kitchen' placed adjacent to a bedroom, or a

Watch

Extended reading notes

Core claim

The central claim is that the mutual information between a future observation and all past observations can be computed on the current scene graph alone, provided one can sample complete scene graphs consistent with what has been seen. The paper treats the scene graph G_k as a sufficient statistic for past observations, so that I(Y_{k+1}, y_{1:k}|x) = I(Y_{k+1}, G_k|x). It samples complete scene graphs from an LLM, defines a future observation as the set of visible, unoccluded nodes in a scene graph from a given viewpoint (with small spatial perturbations to model uncertainty), and uses the resulting predictive distributions to compute an information gain for candidate waypoints. Experiments

Load-bearing premise

The entire information-gain computation assumes the current scene graph is a sufficient statistic of past observations for predicting the next observation; if the graph omits any detail that actually helps predict what will be seen next (wall color, object orientation, partial geometry), the computed gain is not the true mutual information and viewpoint selection can be misled.

Editorial extensions

If this is right

  • For any fixed path length and time budget, the method achieves higher F1 (object-detection accuracy) and lower graph edit distance (topological fidelity) than geometric frontier exploration and voxel-based semantic exploration in three simulated apartments.
  • The method reliably predicts the location of target rooms (e.g., kitchen, bathroom, second bedroom) before the robot directly observes them, reducing the time and distance needed to locate those rooms.
  • Ablations show that explicitly encoding negative space ('nothing' nodes) and adding structure and door nodes successively improves the semantic plausibility of LLM completions and reduces misplaced object proposals.
  • Because the LLM is not told what room to find, the room-prediction behavior emerges from the information-gain objective alone, indicating that the representation itself carries task-relevant spatial semantics.
  • The pipeline runs on a standard GPU-enabled laptop in simulation, suggesting that the added reasoning cost is compatible with embodied deployment.

Reading between the lines

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

  • The equality I(Y_{k+1}, y_{1:k}|x)=I(Y_{k+1}, G_k|x) is the load-bearing assumption; if the scene graph discards predictive detail (color, texture, exact pose), the computed gain is not the true mutual information. A hierarchical belief over graphs—or a graph that explicitly retains predictive low-level features—would be a natural repair.
  • The method's advantage should scale with the statistical regularity of indoor layouts: in environments where the LLM's priors are wrong (unusual floor plans, adversarial scenes), the approach could underperform geometric exploration. This is a testable prediction the paper does not make.
  • Replacing the hand-defined observation model (visible, unoccluded nodes plus perturbations) with a learned renderer from graphs to raw sensor observations would make the information gain directly comparable across representations and could remove the need for manual rejection of hallucinated completions.
  • The manual rejection step (about 0.08 rejections per meter) suggests a dedicated scene-completion model, fine-tuned on indoor layouts, could improve both reliability and speed compared to a general-purpose LLM—a direction the paper mentions as future work.
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 / 5 minor

Summary. The paper proposes an active semantic perception pipeline for indoor exploration. A hierarchical scene graph (rooms, objects, structure, and 'nothing' nodes) is built from RGB-D and segmentation streams, and an LLM (Gemini-2.5-Pro) is prompted to generate a small ensemble of completed scene graphs for unobserved regions. A viewpoint-scoring objective, Eq. (5), combines per-object and per-room information gain computed under a simplified observation model Y(G, x) with node-location perturbations. A two-level planner (global occupancy grid plus local planner) executes the selected waypoint. The approach is evaluated in three HM3D scenes in Habitat against frontier-based exploration and SSMI, reporting F1 and Graph Edit Distance (GED) versus path length/time, plus a room-finding experiment and an ablation over scene-graph node types.

Significance. If the reported gains hold, the paper demonstrates a timely and practically relevant result: LLM priors, injected through sampled scene-graph completions, can make semantic exploration more efficient than geometric frontier-based or voxel-semantic baselines in simulated indoor environments. The manuscript is also refreshingly honest in its limitations section, and the ablation on node types (Nothing, Structure, Door) provides useful qualitative insight. However, the central claim is currently established only for a human-assisted version of the pipeline (manual rejection of LLM outputs), and several reporting inconsistencies (missing SSMI numbers, an unsupported real-robot claim in the abstract) need to be resolved before the contribution can be accepted as stated.

major comments (4)
  1. [Section V, Limitations] The manuscript admits that 'we manually rejected some outputs of the LLM (about 0.08 rejections per meter averaged across the three scenes; with at most two rejections allowed per step).' This is a human-in-the-loop filter that is not part of the algorithm in Section III, is not applied to the baselines, and is not ablated. Because the filter removes implausible completions before they influence waypoint selection, the comparison in Section IV-B conflates the method's algorithmic performance with operator judgment. The claimed superiority 'for any fixed path length and time budget' is therefore not established for the autonomous system. The authors should either run the full pipeline without manual rejection or provide an automated plausibility check and report both variants.
  2. [Eq. (1), Section III-B] The first equality, I(Y_{k+1}, y_{1:k} | x) = I(Y_{k+1}, G_k | x), is asserted to follow because G_k is 'a representation of the past observations.' A representation is not generally a sufficient statistic, and the paper gives no argument that the scene graph captures all information in past observations relevant to predicting Y_{k+1}. This is load-bearing: if G_k discards predictive details (e.g., photometric or fine-grained geometric cues), the computed quantity is not the true information gain. Similarly, the observation model Y(G, x) — visible unoccluded nodes with uncertainty via node-location perturbations — is a surrogate never validated against actual camera observations. Please either justify the sufficiency assumption in a specific setting or explicitly frame the objective as an approximation/heuristic, and add a validation of the observation model against a real sensor model.
  3. [Section IV, Table I and Fig. 3] The experiments section states that the method is compared against two baselines, frontier-based exploration and SSMI, and the Fig. 3 caption mentions SSMI's mean travel distance. However, Table I reports F1/GED for only 'Frontier' and 'Semantic Exploration' — no SSMI entries appear in Table I, and the text does not state whether SSMI is present in Fig. 3. This makes the quantitative comparison to SSMI unclear. The authors should either include SSMI results in Table I and Fig. 3, or explicitly state that SSMI was excluded from the F1/GED comparison and explain why.
  4. [Abstract and Section IV] The abstract claims evaluation 'also on a Unitree Go 2 robot in the real world,' but the body contains no real-world experiments; all reported results are in Habitat simulation (Section IV-A). This unsupported claim should be removed from the abstract or substantiated with actual real-robot data. As written, it overstates the scope of the evaluation.
minor comments (5)
  1. [Section III-B] The notation for mutual information in Eq. (1) uses a comma between arguments; standard notation is a semicolon (I(Y_{k+1} ; y_{1:k} | x)). Also, Eq. (2) writes dp(G_k) but the surrounding text defines p(G|G_k) as a mixture over completed graphs; please clarify the conditioning and the measure being integrated.
  2. [Section IV-A] The setup says 'We build 2 scene graphs using the mapping pipeline to represent p(G_k) in Eq. (2)' — p(G_k) is not defined; presumably p(G|G_k) is intended. Minor typo: 'a subset of observation' should be 'observations.'
  3. [Fig. 3 caption] The caption uses 'Semantic Exploration' to refer to the proposed method, but 'semantic exploration' is also a generic descriptor of SSMI. Rename the proposed method (e.g., 'Ours' or 'SceneGraph-LLM') in the caption and the table to avoid confusion.
  4. [Section IV-C, Table II] In Table II, the Frontier row for scene 00573, Bathroom, reports Tfind/Dfind as '48 5' without a decimal separator or comma. Also, the room-finding comparison includes only the frontier baseline; please state why SSMI is excluded from this experiment beyond the planner difference.
  5. [References] Reference [1] contains a typo: 'scence semantics' should be 'scene semantics.'

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation; the planning objective and evaluation metrics are distinct, and the self-citations are not load-bearing.

full rationale

The paper's derivation chain is not circular. The information-gain objective in Eq. (1)-(5) is computed from LLM-sampled scene-graph completions (Eq. (3)) and node-perturbation renderings (Eq. (4)); no free parameter is fitted to the F1/GED or room-finding metrics (lambda=1 is hand-set, and m=4, N=4 are fixed constants). The F1/GED evaluation in Sec. IV-B measures the constructed scene graph against a manually collected ground-truth graph, not against the LLM completions that drive waypoint selection, so the computed information gain and the reported outcomes are not the same quantity by construction. The room-finding experiment in Sec. IV-C defines a successful prediction as majority-consensus LLM output within 2.5 m of the ground-truth room; this is an operational success criterion, not a prediction defined in terms of the target. Self-citations such as [6] and [19] provide background for the standard mutual-information objective and information-driven exploration; they do not force the present result, and no uniqueness theorem or ansatz is imported from author-only prior work. Two non-circular validity concerns are flagged: Sec. V admits manual rejection of LLM outputs ('we manually rejected some outputs of the LLM ... about 0.08 rejections per meter averaged across the three scenes'), a human-in-the-loop filter not applied to the baselines; and the abstract's claim of evaluation 'also on a Unitree Go 2 robot in the real world' is not supported by any experiment in the body. Eq. (1)'s first equality also assumes G_k is a sufficient statistic of past observations, an unvalidated modeling assumption rather than a circular reduction. These issues affect experimental validity, not circularity, so the circularity score is 0.

Assumptions & free parameters 4 free parameters · 4 assumptions · 1 invented entities

The paper's 'derivation' of information gain is a heuristic build on domain assumptions about the scene graph representation and the LLM's posterior. Many hyperparameters (lambda, sample counts, viewpoint counts) are hand-set. The main invented entity is the 'nothing' node, which is a representation choice rather than a physical postulate.

free parameters (4)
  • lambda (λ) in Eq. (5) = 1
    Weight balancing object-node and room-node information gains; chosen by hand, not tuned or justified.
  • Number of LLM completions (m=4) and number of scene graphs (2), total 8 samples = 8
    Sampling hyperparameters for p(G|G_k); the paper admits 8 samples are not sufficient to capture diversity (Sec. V).
  • Perturbation count (N=4) per scene graph = 4
    Number of node-position perturbations used in Eq. (4) to model observation uncertainty; arbitrary.
  • Candidate viewpoint count = 300
    Uniformly sampled candidate viewpoints; choice affects the max over viewpoints.
assumptions (4)
  • domain assumption The scene graph G_k is a sufficient statistic of past observations for future observation Y_{k+1}.
    Eq. (1) sets I(Y_{k+1}; y_{1:k}|x) = I(Y_{k+1}; G_k|x) without proof; generally false if the graph discards predictive information.
  • domain assumption LLM samples {G^(i)} approximate the posterior p(G|G_k) over plausible scene completions.
    The LLM is not trained for scene expansion; the paper itself notes the proposal distribution can diverge from the true scene distribution (Sec. V).
  • domain assumption Observation Y(G,x) — visible, unoccluded nodes — plus node-location perturbations captures the uncertainty in future observations.
    The rendering model in Sec. III-B treats observations as deterministic node sets; no validation against real RGB-D observations is given.
  • ad hoc to paper Manual rejection of ~8% of LLM outputs does not bias the evaluation.
    Authors manually discarded hallucinated outputs during experiments (Sec. V); this human-in-the-loop step is not part of the algorithm and could inflate reported performance.
invented entities (1)
  • "Nothing" nodes
    purpose: Represent confirmed free space to constrain LLM scene completions and reduce incorrect object placements.
    A new representational primitive introduced in this paper; the ablation suggests it helps, but no external falsifiable prediction is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Active Semantic Perception." pith.science (2026). https://pith.science/paper/2ETSQEY7

@misc{pith2026251005430,
  author       = {Pith},
  title        = {Pith review of: Active Semantic Perception},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2ETSQEY7}},
  note         = {Machine review of arXiv:2510.05430}
}
read the original abstract

We develop an approach for active semantic perception, which refers to using the semantics of the scene for tasks such as exploration. We build a compact, multi-layer scene graph that can represent large, complex indoor environments at various levels of abstraction, e.g., nodes corresponding to rooms, objects, walls, windows etc., as well as fine-grained details of their geometry. We develop a procedure based on large language models (LLMs) to sample new plausible scene graphs of unobserved regions that are consistent with partial observations of the scene. We develop a procedure to compute the information gain of a potential waypoint upon this scene graph to enable sophisticated spatial reasoning: for example, of the two doors that lead out of the living room, one probably leads to the kitchen and the other to the bedroom. We evaluate our approach in realistic 3D indoor apartments in simulation and also on a Unitree Go 2 robot in the real world. Qualitative and quantitative analysis shows that our approach can pin down high-level and low-level semantic information in the environment quickly and more accurately than existing approaches.

Figures

Figures reproduced from arXiv: 2510.05430 by the authors.

Figure 1
Figure 1. Top (left): A robot exploring an indoor environment using our proposed method. The first 15 exploration steps are highlighted in red, while the remaining steps shown in gray. The robot quickly covers most semantically salient regions in spite of visual occlusions and narrow corridors/doors, after which it explores the semantic details of the scene. Top (right): The final scene graph and mesh constructed after the ro… view at source ↗
Figure 2
Figure 2. The LLM produces a semantically plausible completion of a given scene graph that is also roughly consistent in terms of the geometry, e.g., the bed and nightstand are within the bedroom and next to the wardrobe. There is also some degree semantic inconsistency in the LLM completions, e.g., the bed is unlikely to be at this angle right next to the wall. The red part of the scene graph indicates what the camera expect… view at source ↗
Figure 3
Figure 3. Average F1 score and GED as a function of path length [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Visualization of the paths taken to locate a kitchen: semantic exploration vs. baseline. Semantic exploration follows a shorter trajectory and successfully predicts the kitchen’s location before it is observed. outperforms the baseline and it can reliably predict poten…
Figure 5
Figure 5. Figure 5: Ablation on scene-graph node types. Constraints are enabled progressively: (A) Baseline; (B) +Nothing (free-space); (C) +Nothing+Structure; (D) +Nothing+Structure+Door. Step t Prediction Step t+1 Prediction Step t+1 [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Left: The LLM successfully predicts a bathroom above the bedroom for the robot to navigate to. Right: LLM incorrectly predicts a bedroom in the bottom-left corner, misleading the agent to navigate toward an already observed empty region. semantics are stored at the vox…
Figure 7
Figure 7. Figure 7: A summarized version of the generation and refinement prompts used in all experiments. REFERENCES [1] S. He, “Mathematical theory and algorithms for scence semantics in robotics,” B.S. Thesis, University of Pennsylvania, 2025. [2] J. Kirby, An Invitation to Model Theor…

Discussion (0). Sign in to comment.

Forward citations

Cited by 3 Pith papers

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

  1. RGB-only Active 3D Scene Graph Generation for Indoor Mobile Robots

    cs.RO 2026-05 unverdicted novelty 6.0 of 10

    RGB-only active 3D scene graph generation unifies perception and planning to achieve depth-baseline parity and more than double object detection in active indoor exploration.

  2. Fixed External Cameras as Common Prior Maps for Active 3D Scene Graph Generation

    cs.RO 2026-05 unverdicted novelty 6.0 of 10

    Fixed external cameras as Common Prior Maps boost initial object recall in 3D scene graph generation by up to 79% and improve active exploration efficiency.

  3. Bridging Values and Behavior: A Hierarchical Framework for Proactive Embodied Agents

    cs.AI 2026-04 unverdicted novelty 5.0 of 10

    ValuePlanner is a hierarchical architecture that uses LLMs to generate value-based subgoals and PDDL planners to produce executable actions, enabling self-directed behavior in embodied agents.

Reference graph

Works this paper leans on

32 extracted references · 3 linked inside Pith · cited by 3 Pith papers

  1. [1]

    Mathematical theory and algorithms for scence semantics in robotics,

    S. He, “Mathematical theory and algorithms for scence semantics in robotics,” B.S. Thesis, University of Pennsylvania, 2025

    2025

  2. [2]

    Kirby,An Invitation to Model Theory, 2019

    J. Kirby,An Invitation to Model Theory, 2019

    2019

  3. [3]

    Assigning meanings to programs,

    R. W. Floyd, “Assigning meanings to programs,” inProgram Verifica- tion: Fundamental Issues in Computer Science, 1993, pp. 65–81

    1993

  4. [4]

    Clio: Real-time task-driven open-set 3d scene graphs,

    D. Maggio, Y . Chang, N. Hughes, M. Trang, D. Griffith, C. Dougherty, E. Cristofalo, L. Schmid, and L. Carlone, “Clio: Real-time task-driven open-set 3d scene graphs,”IEEE RA-L, 2024

    2024

  5. [5]

    Revisiting active perception,

    R. Bajcsy, Y . Aloimonos, and J. K. Tsotsos, “Revisiting active perception,”Autonomous Robots, vol. 42, no. 2, pp. 177–196, 2018

    2018

  6. [6]

    Active perception using neural radiance fields,

    H. Siming, C. D. Hsu, D. Ong, Y . S. Shao, and P. Chaudhari, “Active perception using neural radiance fields,” inACC, 2024, pp. 4353–4358. 3Prompts have a large effect on LLM outputs. Therefore, we provide a concise summary of the prompts used across all experiments in Fig. 7

    2024

  7. [7]

    Hydra: A real-time spatial perception system for 3D scene graph construction and optimization,

    N. Hughes, Y . Chang, and L. Carlone, “Hydra: A real-time spatial perception system for 3D scene graph construction and optimization,” 2022

    2022

  8. [8]

    Octomap: An efficient probabilistic 3d mapping framework based on octrees,

    A. Hornung, K. M. Wurm, M. Bennewitz, C. Stachniss, and W. Burgard, “Octomap: An efficient probabilistic 3d mapping framework based on octrees,”Autonomous robots, vol. 34, no. 3, pp. 189–206, 2013

    2013

Show all 32 references
  1. [9]

    V oxblox: Incremental 3d euclidean signed distance fields for on-board mav planning,

    H. Oleynikova, Z. Taylor, M. Fehr, R. Siegwart, and J. Nieto, “V oxblox: Incremental 3d euclidean signed distance fields for on-board mav planning,” inIROS, 2017, pp. 1366–1373

    2017

  2. [10]

    Semantic octree mapping and shannon mutual information computation for robot exploration,

    A. Asgharivaskasi and N. Atanasov, “Semantic octree mapping and shannon mutual information computation for robot exploration,”T-RO, vol. 39, no. 3, pp. 1910–1928, 2023

    1910

  3. [11]

    Nerf: Representing scenes as neural radiance fields for view synthesis,

    B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron, R. Ramamoor- thi, and R. Ng, “Nerf: Representing scenes as neural radiance fields for view synthesis,”CACM, vol. 65, no. 1, pp. 99–106, 2021

    2021

  4. [12]

    3d gaussian splatting for real-time radiance field rendering

    B. Kerbl, G. Kopanas, T. Leimk ¨uhler, and G. Drettakis, “3d gaussian splatting for real-time radiance field rendering.”ACM Trans. Graph., vol. 42, no. 4, pp. 139–1, 2023

    2023

  5. [13]

    In-place scene labelling and understanding with implicit scene representation,

    S. Zhi, T. Laidlow, S. Leutenegger, and A. J. Davison, “In-place scene labelling and understanding with implicit scene representation,” in ICCV, 2021, pp. 15 838–15 847

    2021

  6. [14]

    Next best sense: Guiding vision and touch with fisherrf for 3d gaussian splatting,

    M. Strong, B. Lei, A. Swann, W. Jiang, K. Daniilidis, and M. Kennedy, “Next best sense: Guiding vision and touch with fisherrf for 3d gaussian splatting,” inICRA, 2025

    2025

  7. [15]

    A frontier-based approach for autonomous exploration,

    B. Yamauchi, “A frontier-based approach for autonomous exploration,” inCIRA, 1997, pp. 146–151

    1997

  8. [16]

    Learning to explore indoor environments using autonomous micro aerial vehicles,

    Y . Tao, E. Iceland, B. Li, E. Zwecher, U. Heinemann, A. Cohen, A. Avni, O. Gal, A. Barel, and V . Kumar, “Learning to explore indoor environments using autonomous micro aerial vehicles,” inICRA, 2024

    2024

  9. [17]

    Receding horizon

    A. Bircher, M. Kamel, K. Alexis, H. Oleynikova, and R. Siegwart, “Receding horizon” next-best-view” planner for 3d exploration,” in ICRA, 2016, pp. 1462–1468

    2016

  10. [18]

    Information-theoretic mapping using cauchy-schwarz quadratic mutual information,

    B. Charrow, S. Liu, V . Kumar, and N. Michael, “Information-theoretic mapping using cauchy-schwarz quadratic mutual information,” inICRA, 2015, pp. 4791–4798

    2015

  11. [19]

    An active perception game for robust information gathering,

    S. He, Y . Tao, I. Spasojevic, V . Kumar, and P. Chaudhari, “An active perception game for robust information gathering,” inICRA, 2025, pp. 14 168–14 174

    2025

  12. [20]

    Fisherrf: Active view selection and uncertainty quantification for radiance fields using fisher information,

    W. Jiang, B. Lei, and K. Daniilidis, “Fisherrf: Active view selection and uncertainty quantification for radiance fields using fisher information,” inICCV, 2024

    2024

  13. [21]

    Object goal navigation using goal-oriented semantic exploration,

    D. S. Chaplot, D. P. Gandhi, A. Gupta, and R. R. Salakhutdinov, “Object goal navigation using goal-oriented semantic exploration,”NeurIPS, vol. 33, pp. 4247–4258, 2020

    2020

  14. [22]

    How to not train your dragon: Training-free embodied object goal navigation with semantic frontiers,

    J. Chen, G. Li, S. Kumar, B. Ghanem, and F. Yu, “How to not train your dragon: Training-free embodied object goal navigation with semantic frontiers,”RSS, 2023

    2023

  15. [23]

    Yoloe: Real-time seeing anything,

    A. Wang, L. Liu, H. Chen, Z. Lin, J. Han, and G. Ding, “Yoloe: Real-time seeing anything,”arXiv preprint arXiv:2503.07465, 2025

    2025

  16. [24]

    Fast segment anything,

    X. Zhao, W. Ding, Y . An, Y . Du, T. Yu, M. Li, M. Tang, and J. Wang, “Fast segment anything,”arXiv preprint arXiv:2306.12156, 2023

    2023 arXiv

  17. [25]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark,et al., “Learning transferable visual models from natural language supervision,” inICML, 2021, pp. 8748–8763

    2021

  18. [26]

    Marching cubes: A high resolution 3d surface construction algorithm,

    W. E. Lorensen and H. E. Cline, “Marching cubes: A high resolution 3d surface construction algorithm,” inSeminal graphics, 1998, pp. 347–353

    1998

  19. [27]

    vs-graphs: Integrating visual slam and situational graphs through multi-level scene understanding,

    A. Tourani, S. Ejaz, H. Bavle, D. Morilla-Cabello, J. L. Sanchez- Lopez, and H. V oos, “vs-graphs: Integrating visual slam and situational graphs through multi-level scene understanding,”arXiv preprint arXiv:2503.01783, 2025

    2025 arXiv

  20. [28]

    You only segment once: Towards real-time panoptic segmentation,

    J. Hu, L. Huang, T. Ren, S. Zhang, R. Ji, and L. Cao, “You only segment once: Towards real-time panoptic segmentation,” inCVPR, 2023, pp. 17 819–17 829

    2023

  21. [29]

    Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography,

    M. A. Fischler and R. C. Bolles, “Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography,”CACM, vol. 24, no. 6, pp. 381–395, 1981

    1981

  22. [30]

    Habitat-matterport 3d dataset (HM3d): 1000 large-scale 3d environments for embodied AI,

    S. K. Ramakrishnan, A. Gokaslan, E. Wijmans, O. Maksymets, A. Clegg, J. M. Turner, E. Undersander, W. Galuba, A. Westbury, A. X. Chang, M. Savva, Y . Zhao, and D. Batra, “Habitat-matterport 3d dataset (HM3d): 1000 large-scale 3d environments for embodied AI,” inNeurIPS Dataset...

    2021

  23. [31]

    Habitat: A platform for embodied ai research,

    M. Savva, A. Kadian, O. Maksymets, Y . Zhao, E. Wijmans, B. Jain, J. Straub, J. Liu, V . Koltun, J. Malik,et al., “Habitat: A platform for embodied ai research,” inICCV, 2019, pp. 9339–9347

    2019

  24. [32]

    Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities,

    G. Comanici, E. Bieber, M. Schaekermann, I. Pasupat, N. Sachdeva, I. Dhillon, M. Blistein, O. Ram, D. Zhang, E. Rosen,et al., “Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities,”arXiv preprint arXiv:...

    2025 arXiv

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.