Pith. sign in

REVIEW 3 major objections 5 minor 3 cited by

VISTA: Open-Vocabulary, Task-Relevant Robot Exploration with Online Semantic Gaussian Splatting

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read VISTA claims that fusing geometric view-diversity with semantic similarity lets a quadrotor find occluded query objects in all six hard-map trials, a sixfold success-rate improvement over its baselines.

desk verdict VISTA's view-diversity metric is a genuine contribution and the static NBV results are convincing, but the headline 6x hardware success claim rests on an undefined stopping condition and tiny samples. read the letter →

arxiv 2507.01125 v1 pith:ADAADP4L submitted 2025-07-01 cs.RO

Keiko Nagami , Timothy Chen , Javier Yu , Ola Shorinwa , Maximilian Adang , Carlyn Dougherty , Eric Cristofalo , Mac Schwager This is my paper
classification cs.RO
keywords activeexplorationsemantic3DGaussiansplattingopen-vocabularyobjectsearchnext-best-viewselectioninformationgainvision-languagemodelsquadrotorplanningreceding-horizon
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

VISTA is trying to establish a simple rule for task-driven exploration: when a robot is told to find an object by name, it should plan toward viewpoints that both raise the semantic match to the query and reduce geometric uncertainty, and a cheap angular-diversity measure can serve as that geometric uncertainty online. The system builds a live semantic 3D Gaussian Splatting map, distills CLIP embeddings into it, and scores sampled trajectories with a weighted geometric-plus-semantic gain. On the paper's evidence, this rule lets a quadrotor succeed in all six occluded-object trials while semantic-only and geometric-only baselines each succeed in one, and the geometric metric also selects better next views than two prior radiance-field baselines. If true, open-vocabulary object search in unmapped environments no longer needs a pre-built concept graph or a trained object detector.

What carries the argument

The load-bearing mechanism is the per-voxel viewing-direction history and the coverage gain $g_I(d_x^n)=\frac{1}{2}\,(1-\min_v d_v^\top d_x^n)$, where $d_x^n$ is a ray from the candidate camera pose and $d_v$ are the directions from which that voxel has already been seen; the minimum dot product picks the closest prior view, so a novel angle scores high and a redundant angle scores low. Rays that terminate in unobserved voxels receive the maximum gain. The per-image geometric gain is the mean of these pixel values, added to the mean semantic similarity, and accumulated over a sampled waypoint path with a discount factor to form the VISTA-Score used to pick the next trajectory. Candidate paths are drawn from a Gaussian mixture over frontier cells and high-semantic cells, then executed in a receding-horizon loop with the geometric weight decaying so semantic pull grows over time.

What would settle it

Re-run the six hard-map quadrotor trials exactly as described, but decide success by an independent check — a human watching the live or rendered view, or a separately trained detector — instead of VISTA's own semantic heatmap. If the independently verified success rate falls below 100%, the headline 6x improvement depends on the unstated stopping criterion.

Watch

Extended reading notes

Core claim

The paper's central claim is that geometric view diversity and task semantics belong in the same exploration objective, and that combining them is what makes object search succeed when the target starts out occluded. VISTA represents the scene as a semantic 3D Gaussian Splatting map, flattens it into a voxel grid with per-voxel viewing-direction histories, and scores receding-horizon trajectories by a summed discounted gain of geometric ray-angle coverage plus CLIP semantic similarity. In the quadrotor experiments this leads to 100% success on all six hard-map trials versus 16.67% for both baselines, and on six static scenes the geometric component reaches higher PSNR and SSIM and lower LPIPS than the two prior next-best-view baselines. The authors also report the system running on a quadruped, supporting the claim that the method is platform-agnostic.

Load-bearing premise

The load-bearing premise is that the hardware trials' 'success' labels mean the robot actually found the queried object, but the paper never defines the stopping condition, detector, or ground-truth check used to declare success.

Editorial extensions

If this is right

  • A robot running VISTA can keep exploring after the query object is hidden, because geometric gain keeps rewarding viewpoints that reveal unobserved space rather than converging on the current best semantic spot.
  • The same ray-angle coverage metric works as a fast next-best-view selector for photorealistic mapping, reaching comparable image quality with fewer training iterations than Fisher information or Bayesian uncertainty baselines.
  • Open-vocabulary queries need no per-object detector: any CLIP-addressable phrase can drive the search, and the same pipeline transfers from a quadrotor to a quadruped.
  • Because the map and the gain metric are updated online, the robot interleaves mapping and searching instead of following the two-stage map-then-search pattern common in pre-mapped object localization.

Reading between the lines

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

  • Beyond the paper, the reported 6x improvement rests on an undefined success criterion: Section V-B never states the stopping-condition trigger or a detector that verified the object was found, so an independent human or detector-based success label is the natural next check.
  • The view-diversity principle is not tied to Gaussian Splatting; the paper notes a NeRF backend would work, so replacing the 3DGS backend while keeping the voxel-direction history would directly test how much of the gain comes from the metric versus the representation.
  • The decaying geometric weight suggests the planner's behavior is time-sensitive; measuring whether success and time-to-reach degrade when the decay rate is changed would map how much exploration is actually needed versus semantic greed.
  • CLIP's object-centric semantics limit VISTA to object search, but the planning objective itself should extend to richer language embeddings, so a natural stress test is a query like 'find the red cone near the door' where spatial relations matter.
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

3 major / 5 minor

Summary. The paper presents VISTA, a robot exploration system that builds an online semantic 3D Gaussian Splatting map and plans receding-horizon trajectories using a combined geometric and semantic information-gain metric. The geometric metric scores candidate views by angular diversity relative to previously observed directions, and the semantic term uses CLIP embeddings distilled into the map. The authors evaluate the geometric metric on six static scenes against FisherRF and Bayes' Rays using PSNR, SSIM, and LPIPS, report quadrotor hardware experiments in easy and hard occlusion maps against two baselines, and demonstrate the system on a Spot quadruped. The central claim is a 6x improvement in success rate on hard maps relative to baselines.

Significance. If the results hold, VISTA is a useful contribution to active mapping and open-vocabulary object search. The static next-best-view evaluation is a meaningful, externally grounded check of the geometric metric: it is measured against held-out photometric quality, not against VISTA's own objective, and the reported improvements over FisherRF and Bayes' Rays across six scenes give credible evidence that the angular-diversity gain is informative. The real-time capability on two very different robot platforms is also a strength, as is the plan to release code. However, the headline hardware success-rate claim is not decidable from the paper as written, because the success criterion is never defined and the number of trials per condition is small. The static results support the geometric component, but the full system claim needs a more rigorous experimental protocol before it can be accepted.

major comments (3)
  1. [Section V-B, Table I] The success criterion for the quadrotor hardware experiments is never specified. The text says that in the easy map the object should be 'relatively easy to find and have the stopping condition trigger,' but no detector, threshold, human-annotation protocol, or external ground-truth check is described. Given that the planner optimizes a CLIP-based semantic similarity field, if success is declared when that same similarity map exceeds a threshold, the 100% hard-map success rate could measure the planner's confidence in its own map rather than actual discovery of the query object. This is load-bearing because the abstract's '6x higher success rates' is computed from Table I. Please define the stopping condition precisely and, ideally, verify each successful trial independently (e.g., by a human labeler or a separate object detector).
  2. [Section V-B, Table I] The hardware evaluation uses only two trials per query-map cell, totaling 12 trials per method, with no variance, confidence intervals, or significance testing. For the hard-map comparison, VISTA succeeds in 6/6 trials and each baseline in 1/6 trials; a single relabeled trial would change the reported factor from 6x to 3x or less. The paper should report per-trial outcomes, provide a statistical measure (e.g., exact binomial confidence intervals), or otherwise justify that the difference is not attributable to small-sample noise. At minimum, the number of trials and the definition of each successful trial must be transparent.
  3. [Section V-B, paragraph 2] The geometric baseline is said to be 'based off the work in RT-Guide' and is described as an adaptation that seeds the GMM with high-uncertainty Gaussians and scores trajectories by the counts of high- and low-uncertainty Gaussians. If this is not the original RT-GuIDE planner, the comparison may not reflect the performance of RT-GuIDE as published. Please clarify whether the original RT-GuIDE method was run, whether the authors reimplemented it, and what parameters or architectural changes were made; otherwise the claim of outperforming RT-GuIDE is not fully supported.
minor comments (5)
  1. [Section IV-C, Eq. (2)] The notation in Eq. (2) is slightly ambiguous: the min is over the existing direction vectors stored in the voxel, but the indices of the resulting vector are not defined. Please clarify by writing the minimization explicitly with respect to the set of stored directions.
  2. [Section IV-B] There is a typo in the first paragraph: 'V oxel' should be 'Voxel.' Also, the sentence 'Voxel traversal of the voxel map can similarly be used...' has an extra space.
  3. [Figure 3, caption] The caption says 'for about 50K iterations in three visual reconstruction metrics and computation time.' It is not clear what 'about 50K iterations' refers to, since the text says models are trained for 1000 iterations per step. Please rephrase to describe the evaluation protocol consistently.
  4. [Section V-B, Table I] In Table I, rows with zero successes report TTR as 'N/A' and SPL as 0. That is reasonable, but the table would be clearer if it also indicated the time limit imposed by battery life, since TTR is only defined for successful trials within that limit.
  5. [Section IV-D] Several planner hyperparameters (geometric weight c, decay factor beta, discount factor gamma, top-m semantic cells, voxel grid size and resolution, number of sampled trajectories) are introduced but not given values or a sensitivity study. Adding a table of the values used in the hardware and static experiments would improve reproducibility.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity established; the geometric metric is validated against external photometric benchmarks, and the hardware success claim is limited by an unspecified stopping condition rather than by a demonstrable reduction to the planner's own objective.

full rationale

The central information-gain claim is not circular. VISTA's geometric metric (Eq. 2-3) is a heuristic view-diversity score, and its static evaluation is measured against external photometric metrics: 'We evaluate each method using the standard metrics: Peak-Signal-Noise-Ratio (PSNR), Learned Perceptuation Image Patch Similarity (LPIPS), and Structural Similarity Index Measure (SSIM)' (Section V-A). Those targets are independent of VISTA's own score, so the next-best-view comparison is self-contained. The hardware experiment compares whole-system success rates, not a derived quantity, so no equation-level equivalence to inputs is present. The main concern is the undefined stopping condition: Section V-B says 'the query object should be relatively easy to find and have the stopping condition trigger' but never defines the detector, threshold, human-annotation rule, or external ground-truth check, and Section III frames the task as determining 'whether or not the identified object satisfies the user's input query.' If the stopping trigger were implemented from VISTA's own CLIP-based semantic field, then success would be measured by the same semantic similarity that the planner maximizes in Eq. (4), which would be a real circularity. However, the paper does not say this, so under the hard rule requiring quoted evidence of a reduction, this remains an evaluation-protocol gap rather than an established circular step. The self-citations to NerfBridge [48] and Splatbridge/Safer-Splat [50] by overlapping authors are used as implementation infrastructure for online radiance-field training, not as evidence for the information-gain or success-rate claims, so they are not load-bearing. No uniqueness theorem is imported, and no fitted parameter is renamed as a prediction. Overall, the derivations are self-contained; the score of 2 reflects the unresolved stopping-condition ambiguity rather than a demonstrated circular derivation.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central planning objective depends on several unreported hyperparameters (c, beta, gamma, m, grid resolution) and on heuristics (view-diversity proxy, GMM sampling) that are not derived from first principles. No new physical entities are introduced.

free parameters (6)
  • geometric weight c = not reported
    Scales GI(x) in Eq (4) and is decayed by beta^i in Algorithm 1, line 14; its trajectory strongly influences whether the robot explores geometrically or chases semantics.
  • decay factor beta = not reported
    Controls how quickly the geometric weight c decays over replanning index i (Algorithm 1, line 14); no value or schedule is given.
  • discount factor gamma = not reported
    Discounts future waypoints in the trajectory score, Eq (4); no value is given.
  • top-m semantic cells = not reported
    Number of highest-semantic 2D cells used to build the categorical distribution Ds in Algorithm 1, line 3; affects GMM sampling bias.
  • voxel grid size and resolution = not reported
    Section IV-B fixes grid size and resolution centered on the robot, but the values are not stated; the coverage metric and planning depend on this discretization.
  • candidate trajectory count and GMM components = not reported
    Algorithm 1, line 5 samples a set of trajectories from a GMM; the number of components and samples are unspecified, affecting planner behavior.
assumptions (5)
  • domain assumption View-direction diversity is a strong proxy for geometric reconstruction quality.
    Stated as the key insight in the Introduction; the entire geometric gain gI (Eq. 2) is built on it. Indirectly supported by the six-scene NBV comparison, but not proven in general.
  • domain assumption CLIP image embeddings distilled into 3D points encode task relevance for open-vocabulary queries.
    Section IV-A trains a semantic field with CLIP embeddings; the semantic gain assumes these similarities predict object presence. The authors note CLIP is object-centric in the Limitations.
  • domain assumption The robot has reliable full-pose state estimation.
    Section III assumes full pose odometry; experiments use an OptiTrack motion-capture system or onboard SLAM. Pose errors would corrupt 3DGS training, voxel traversal, and planning.
  • domain assumption Robot motion can be modeled as a planar single integrator with yaw control.
    Section III restricts motion to fixed z, phi, theta; valid for the tested platforms but excludes 3D flight or uneven terrain maneuvers.
  • ad hoc to paper GMM sampling over frontier cells and top semantic cells yields a representative set of candidate trajectories.
    Algorithm 1 lines 2-5 fit a GMM to Df and Ds and sample trajectories; no coverage or optimality guarantee is provided for this proposal distribution.

how reviews work

0 comments
Cite this review

Pith. "Pith review of VISTA: Open-Vocabulary, Task-Relevant Robot Exploration with Online Semantic Gaussian Splatting." pith.science (2026). https://pith.science/paper/ADAADP4L

@misc{pith2026250701125,
  author       = {Pith},
  title        = {Pith review of: VISTA: Open-Vocabulary, Task-Relevant Robot Exploration with Online Semantic Gaussian Splatting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ADAADP4L}},
  note         = {Machine review of arXiv:2507.01125}
}
read the original abstract

We present VISTA (Viewpoint-based Image selection with Semantic Task Awareness), an active exploration method for robots to plan informative trajectories that improve 3D map quality in areas most relevant for task completion. Given an open-vocabulary search instruction (e.g., "find a person"), VISTA enables a robot to explore its environment to search for the object of interest, while simultaneously building a real-time semantic 3D Gaussian Splatting reconstruction of the scene. The robot navigates its environment by planning receding-horizon trajectories that prioritize semantic similarity to the query and exploration of unseen regions of the environment. To evaluate trajectories, VISTA introduces a novel, efficient viewpoint-semantic coverage metric that quantifies both the geometric view diversity and task relevance in the 3D scene. On static datasets, our coverage metric outperforms state-of-the-art baselines, FisherRF and Bayes' Rays, in computation speed and reconstruction quality. In quadrotor hardware experiments, VISTA achieves 6x higher success rates in challenging maps, compared to baseline methods, while matching baseline performance in less challenging maps. Lastly, we show that VISTA is platform-agnostic by deploying it on a quadrotor drone and a Spot quadruped robot. Open-source code will be released upon acceptance of the paper.

Figures

Figures reproduced from arXiv: 2507.01125 by the authors.

Figure 1
Figure 1. System overview of VISTA. Real-time sensor data is gathered from [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Geometric information gain based on view diversity coverage. Given [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Our geometric information gain metric significantly outperforms baselines FisherRF and Bayes Rays in the next best view selection task for about [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: The top row shows our three environments and two robots, with the search object in a green circle. The second row shows an example trajectory [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

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. Coverage Optimization for Camera View Selection

    cs.CV 2026-04 unverdicted novelty 6.0 of 10

    COVER is a new coverage metric that selects camera views by prioritizing insufficiently observed geometry, yielding better NeRF reconstructions than prior active selection methods.

  2. Multi-Agent Next-Best-View Optimization for Risk-Averse Planning

    cs.RO 2026-06 unverdicted novelty 5.0 of 10

    Distributed risk-aware NBV framework for multi-agent planning using private 3DGS maps, C-ADMM optimization, and AV@R risk scoring that approaches centralized performance with orders-of-magnitude less communication.

  3. Semantic-Aware Guided Drone Exploration for Language-Conditioned 3D Indoor Mapping

    cs.RO 2026-05 unverdicted novelty 5.0 of 10

    SAGE integrates CLIP embeddings into FALCON via object storage, temporal projection cache, object frontiers, and a bounded semantic-geometric cost, yielding 9-26x faster object discovery than FTU in simulation and hig...

Reference graph

Works this paper leans on

52 extracted references · 34 canonical work pages · cited by 3 Pith papers

  1. [1]

    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,” in International Conference on Machine Learning (ICML) . PMLR, 2021, pp. 8748– 8763

    2021

  2. [2]

    Emerging properties in self-supervised vision trans- formers,

    M. Caron, H. Touvron, I. Misra, H. J ´egou, J. Mairal, P. Bojanowski, and A. Joulin, “Emerging properties in self-supervised vision trans- formers,” in Proceedings of the IEEE/CVF international conference on computer vision , 2021, pp. 9650–9660

    work page 2021

  3. [3]

    Foundation models in robotics: Applications, challenges, and the future,

    R. Firoozi, J. Tucker, S. Tian, A. Majumdar, J. Sun, W. Liu, Y . Zhu, S. Song, A. Kapoor, K. Hausman et al. , “Foundation models in robotics: Applications, challenges, and the future,” The International Journal of Robotics Research , p. 02783649241281508, 2023

    work page 2023

  4. [4]

    Conceptgraphs: Open-vocabulary 3d scene graphs for perception and planning,

    Q. Gu, A. Kuwajerwala, S. Morin, K. M. Jatavallabhula, B. Sen, A. Agarwal, C. Rivera, W. Paul, K. Ellis, R. Chellappa et al. , “Conceptgraphs: Open-vocabulary 3d scene graphs for perception and planning,” in 2024 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2024, pp. 5021–5028

    2024

  5. [5]

    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

  6. [6]

    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,” Communications of the ACM , vol. 65, no. 1, pp. 99–106, 2021

    2021

  7. [7]

    Using occupancy grids for mobile robot perception and navigation,

    A. Elfes, “Using occupancy grids for mobile robot perception and navigation,” Computer, vol. 22, no. 6, pp. 46–57, Jun. 1989

    work page 1989

  8. [8]

    Slam-driven robotic mapping and registration of 3d point clouds,

    P. Kim, J. Chen, and Y . K. Cho, “Slam-driven robotic mapping and registration of 3d point clouds,” Automation in Construction , vol. 89, pp. 38–48, 2018

    work page 2018

Show all 52 references
  1. [9]

    3d mapping with multi-resolution occupied voxel lists,

    J. Ryde and H. Hu, “3d mapping with multi-resolution occupied voxel lists,” Autonomous Robots, vol. 28, pp. 169–185, 2010

    2010

  2. [10]

    A frontier-based approach for autonomous exploration,

    B. Yamauchi, “A frontier-based approach for autonomous exploration,” in Proceedings 1997 IEEE International Symposium on Computational Intelligence in Robotics and Automation CIRA’97. ’Towards New Com- putational Principles for Robotics and Automation’ . IEEE, 1997, pp. 146–151

    1997

  3. [11]

    Echo: An efficient heuristic viewpoint determination method on frontier-based autonomous explo- ration for quadrotors,

    J. Yu, H. Shen, J. Xu, and T. Zhang, “Echo: An efficient heuristic viewpoint determination method on frontier-based autonomous explo- ration for quadrotors,” IEEE Robotics and Automation Letters , vol. 8, no. 8, pp. 5047–5054, 2023

    2023

  4. [12]

    Rapid exploration with multi-rotors: A frontier selection method for high speed flight,

    T. Cieslewski, E. Kaufmann, and D. Scaramuzza, “Rapid exploration with multi-rotors: A frontier selection method for high speed flight,” in 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2017, pp. 2135–2142

    2017

  5. [13]

    Receding horizon path planning for 3d exploration and surface inspection,

    A. Bircher, M. Kamel, K. Alexis, H. Oleynikova, and R. Siegwart, “Receding horizon path planning for 3d exploration and surface inspection,” Autonomous Robots, vol. 42, pp. 291–306, 2018

    2018

  6. [14]

    Information theoretic active exploration in signed distance fields,

    K. Saulnier, N. Atanasov, G. J. Pappas, and V . Kumar, “Information theoretic active exploration in signed distance fields,” in 2020 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2020, pp. 4080–4085

    2020

  7. [15]

    Seer: Safe efficient exploration for aerial robots using learning to predict information gain,

    Y . Tao, Y . Wu, B. Li, F. Cladera, A. Zhou, D. Thakur, and V . Kumar, “Seer: Safe efficient exploration for aerial robots using learning to predict information gain,” in 2023 IEEE International Conference on Robotics and Automation (ICRA) . IEEE, 2023, pp. 1235–1241

    2023

  8. [16]

    Informed sampling exploration path planner for 3d reconstruction of large scenes,

    Y . Kompis, L. Bartolomei, R. Mascaro, L. Teixeira, and M. Chli, “Informed sampling exploration path planner for 3d reconstruction of large scenes,” IEEE Robotics and Automation Letters , vol. 6, no. 4, pp. 7893–7900, 2021

    2021

  9. [17]

    Fisherrf: Active view selection and uncertainty quantification for radiance fields using fisher informa- tion,

    W. Jiang, B. Lei, and K. Daniilidis, “Fisherrf: Active view selection and uncertainty quantification for radiance fields using fisher informa- tion,” arXiv preprint arXiv:2311.17874 , 2023

    2023 arXiv

  10. [18]

    Mip-NeRF360: Unbounded anti-aliased neural radiance fields,

    J. T. Barron, B. Mildenhall, D. Verbin, P. P. Srinivasan, and P. Hed- man, “Mip-NeRF360: Unbounded anti-aliased neural radiance fields,” IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022

    2022

  11. [19]

    Instant neural graphics primitives with a multiresolution hash encoding,

    T. M ¨uller, A. Evans, C. Schied, and A. Keller, “Instant neural graphics primitives with a multiresolution hash encoding,” ACM Trans. Graph., vol. 41, no. 4, pp. 102:1–102:15, Jul. 2022

    2022

  12. [20]

    Mip-splatting: Alias-free 3d gaussian splatting,

    Z. Yu, A. Chen, B. Huang, T. Sattler, and A. Geiger, “Mip-splatting: Alias-free 3d gaussian splatting,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2024, pp. 19 447–19 456

    2024

  13. [21]

    NeRF-SLAM: Real-Time Dense Monocular SLAM with Neural Radiance Fields,

    A. Rosinol, J. J. Leonard, and L. Carlone, “NeRF-SLAM: Real-Time Dense Monocular SLAM with Neural Radiance Fields,” arXiv preprint arXiv:2210.13641, 2022

    2022 arXiv

  14. [22]

    Gs- slam: Dense visual slam with 3d gaussian splatting,

    C. Yan, D. Qu, D. Xu, B. Zhao, Z. Wang, D. Wang, and X. Li, “Gs- slam: Dense visual slam with 3d gaussian splatting,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog- nition, 2024, pp. 19 595–19 604

    2024

  15. [23]

    Gaussian splatting slam,

    H. Matsuki, R. Murai, P. H. Kelly, and A. J. Davison, “Gaussian splatting slam,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2024, pp. 18 039–18 048

    2024

  16. [24]

    Activermap: Radiance field for active mapping and planning,

    H. Zhan, J. Zheng, Y . Xu, I. Reid, and H. Rezatofighi, “Activermap: Radiance field for active mapping and planning,” arXiv preprint arXiv:2211.12656, 2022

    2022 arXiv

  17. [25]

    Uncer- tainty guided policy for active robotic 3d reconstruction using neural radiance fields,

    S. Lee, L. Chen, J. Wang, A. Liniger, S. Kumar, and F. Yu, “Uncer- tainty guided policy for active robotic 3d reconstruction using neural radiance fields,” IEEE Robotics and Automation Letters , vol. 7, no. 4, pp. 12 070–12 077, 2022

    2022

  18. [26]

    Active neural mapping,

    Z. Yan, H. Yang, and H. Zha, “Active neural mapping,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 10 981–10 992

    2023

  19. [27]

    Active implicit object reconstruction using uncertainty-guided next-best-view optimization,

    D. Yan, J. Liu, F. Quan, H. Chen, and M. Fu, “Active implicit object reconstruction using uncertainty-guided next-best-view optimization,” IEEE Robotics and Automation Letters , 2023

    2023

  20. [28]

    Gs-planner: A gaussian-splatting- based planning framework for active high-fidelity reconstruction,

    R. Jin, Y . Gao, H. Lu, and F. Gao, “Gs-planner: A gaussian-splatting- based planning framework for active high-fidelity reconstruction,” arXiv preprint arXiv:2405.10142 , 2024

    2024 arXiv

  21. [29]

    Hgs-planner: Hierarchical planning framework for active scene reconstruction using 3d gaussian splatting,

    Z. Xu, R. Jin, K. Wu, Y . Zhao, Z. Zhang, J. Zhao, Z. Gan, and W. Ding, “Hgs-planner: Hierarchical planning framework for active scene reconstruction using 3d gaussian splatting,” arXiv preprint arXiv:2409.17624, 2024

    2024 arXiv

  22. [30]

    Ag-slam: Active gaussian splatting slam,

    W. Jiang, B. Lei, K. Ashton, and K. Daniilidis, “Ag-slam: Active gaussian splatting slam,” arXiv preprint arXiv:2410.17422 , 2024

    2024 arXiv

  23. [31]

    Activegs: Active scene reconstruction using gaussian splatting,

    L. Jin, X. Zhong, Y . Pan, J. Behley, C. Stachniss, and M. Popovi ´c, “Activegs: Active scene reconstruction using gaussian splatting,”IEEE Robotics and Automation Letters , 2025

    2025

  24. [32]

    Beyond uncertainty: Risk-aware active view acquisition for safe robot navigation and 3d scene understanding with fisherrf,

    G. Liu, W. Jiang, B. Lei, V . Pandey, K. Daniilidis, and N. Motee, “Beyond uncertainty: Risk-aware active view acquisition for safe robot navigation and 3d scene understanding with fisherrf,” arXiv preprint arXiv:2403.11396, 2024

    2024 arXiv

  25. [33]

    Bayes’ rays: Uncertainty quantification for neural radiance fields,

    L. Goli, C. Reading, S. Sell ´an, A. Jacobson, and A. Tagliasacchi, “Bayes’ rays: Uncertainty quantification for neural radiance fields,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 20 061–20 070

    2024

  26. [34]

    Rt-guide: Real-time gaussian splatting for information-driven explo- ration,

    Y . Tao, D. Ong, V . Murali, I. Spasojevic, P. Chaudhari, and V . Kumar, “Rt-guide: Real-time gaussian splatting for information-driven explo- ration,” arXiv preprint arXiv:2409.18122 , 2024

    2024

  27. [35]

    LERF: Language embedded radiance fields,

    J. Kerr, C. M. Kim, K. Goldberg, A. Kanazawa, and M. Tancik, “LERF: Language embedded radiance fields,” in IEEE/CVF Inter- national Conference on Computer Vision (ICCV) , 2023, pp. 19 729– 19 739

    2023

  28. [36]

    Feature 3dgs: Supercharging 3d gaussian splatting to enable distilled feature fields,

    S. Zhou, H. Chang, S. Jiang, Z. Fan, Z. Zhu, D. Xu, P. Chari, S. You, Z. Wang, and A. Kadambi, “Feature 3dgs: Supercharging 3d gaussian splatting to enable distilled feature fields,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2024, p...

    2024

  29. [37]

    Langsplat: 3d language gaussian splatting,

    M. Qin, W. Li, J. Zhou, H. Wang, and H. Pfister, “Langsplat: 3d language gaussian splatting,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2024, pp. 20 051–20 060

    2024

  30. [38]

    Fast-splat: Fast, ambiguity- free semantics transfer in gaussian splatting,

    O. Shorinwa, J. Sun, and M. Schwager, “Fast-splat: Fast, ambiguity- free semantics transfer in gaussian splatting,” arXiv preprint arXiv:2411.13753, 2024

    2024 arXiv

  31. [39]

    Hammer: Heterogeneous, multi- robot semantic gaussian splatting,

    J. Yu, T. Chen, and M. Schwager, “Hammer: Heterogeneous, multi- robot semantic gaussian splatting,” arXiv preprint arXiv:2501.14147 , 2025

    2025 arXiv

  32. [40]

    Siren: Se- mantic, initialization-free registration of multi-robot gaussian splatting maps,

    O. Shorinwa, J. Sun, M. Schwager, and A. Majumdar, “Siren: Se- mantic, initialization-free registration of multi-robot gaussian splatting maps,” arXiv preprint arXiv:2502.06519 , 2025

    2025 arXiv

  33. [41]

    Segment anything,

    A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. Whitehead, A. C. Berg, W.-Y . Lo et al. , “Segment anything,” in Proceedings of the IEEE/CVF International Conference on Computer Vision , 2023, pp. 4015–4026

    2023

  34. [42]

    Visual instruction tuning,

    H. Liu, C. Li, Q. Wu, and Y . J. Lee, “Visual instruction tuning,” Advances in neural information processing systems , vol. 36, 2024

    2024

  35. [43]

    Cows on pasture: Baselines and benchmarks for language-driven zero-shot object navigation,

    S. Y . Gadre, M. Wortsman, G. Ilharco, L. Schmidt, and S. Song, “Cows on pasture: Baselines and benchmarks for language-driven zero-shot object navigation,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2023, pp. 23 171–23 181

    2023

  36. [44]

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

    N. Yokoyama, S. Ha, D. Batra, J. Wang, and B. Bucher, “Vlfm: Vision- language frontier maps for zero-shot semantic navigation,” in 2024 IEEE International Conference on Robotics and Automation (ICRA) . IEEE, 2024, pp. 42–48

    2024

  37. [45]

    Atlas navigator: Active task-driven language-embedded gaussian splatting,

    D. Ong, Y . Tao, V . Murali, I. Spasojevic, V . Kumar, and P. Chaudhari, “Atlas navigator: Active task-driven language-embedded gaussian splatting,” 2025. [Online]. Available: https://arxiv.org/abs/2502.20386

    2025 arXiv

  38. [46]

    Finding things in the unknown: Semantic object- centric exploration with an mav,

    S. Papatheodorou, N. Funk, D. Tzoumanikas, C. Choi, B. Xu, and S. Leutenegger, “Finding things in the unknown: Semantic object- centric exploration with an mav,” in 2023 IEEE International Confer- ence on Robotics and Automation (ICRA) . IEEE, 2023, pp. 3339– 3345

    2023

  39. [47]

    Findanything: Open-vocabulary and object-centric mapping for robot exploration in any environment,

    S. Barbas Laina, S. Boche, S. Papatheodorou, S. Schaefer, J. Jung, and S. Leutenegger, “Findanything: Open-vocabulary and object-centric mapping for robot exploration in any environment,” arXiv e-prints , pp. arXiv–2504, 2025

    2025

  40. [48]

    Nerfbridge: Bringing real-time, online neural radiance field training to robotics,

    J. Yu, J. E. Low, K. Nagami, and M. Schwager, “Nerfbridge: Bringing real-time, online neural radiance field training to robotics,” arXiv preprint arXiv:2305.09761, 2023

    2023 arXiv

  41. [49]

    Nerfstudio: A modular framework for neural radiance field development,

    M. Tancik, E. Weber, E. Ng, R. Li, B. Yi, J. Kerr, T. Wang, A. Kristoffersen, J. Austin, K. Salahi, A. Ahuja, D. McAllister, and A. Kanazawa, “Nerfstudio: A modular framework for neural radiance field development,” in ACM SIGGRAPH 2023 Conference Proceedings, ser. SIGGRAPH ’23, 2023

    2023

  42. [50]

    Safer-splat: A control barrier function for safe navigation with online gaussian splatting maps,

    T. Chen, A. Swann, J. Yu, O. Shorinwa, R. Murai, M. Kennedy III, and M. Schwager, “Safer-splat: A control barrier function for safe navigation with online gaussian splatting maps,” arXiv preprint arXiv:2409.09868, 2024

    2024 arXiv

  43. [51]

    Dis- tilled feature fields enable few-shot language-guided manipulation,

    W. Shen, G. Yang, A. Yu, J. Wong, L. P. Kaelbling, and P. Isola, “Dis- tilled feature fields enable few-shot language-guided manipulation,” in 7th Annual Conference on Robot Learning , 2023

    2023

  44. [52]

    A fast voxel traversal algorithm for ray tracing,

    J. Amanatides and A. Woo, “A fast voxel traversal algorithm for ray tracing,” Proceedings of EuroGraphics, vol. 87, 08 1987

    1987

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.