Pith. sign in

REVIEW 3 major objections 4 minor 37 references

SeqAffordSplat: Scene-level Sequential Affordance Reasoning on 3D Gaussian Splatting

T0 review · 3 major / 4 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 The paper claims that sequential, scene-level 3D affordance reasoning can be solved end-to-end by an LLM that autoregressively emits segmentation tokens, each triggering a 3D mask over Gaussian-splat scenes.

desk verdict A genuinely new task and benchmark for sequential 3DGS affordance, with a promising architecture, but the headline SOTA numbers rest on an under-specified evaluation protocol. read the letter →

arxiv 2507.23772 v1 pith:YBH7IIME submitted 2025-07-31 cs.CV

Di Li , Jie Feng , Jiahao Chen , Weisheng Dong , Guanbin Li , Yuhui Zheng , Mingtao Feng , Guangming Shi This is my paper · ORCID
classification cs.CV
keywords 3DaffordancereasoningGaussianSplattingsequentiallargelanguagemodelscene-levelunderstandingmultimodalsegmentationbenchmarkdatasetsemanticfeatureinjection
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

3D affordance reasoning—locating the part of an object that supports a requested action like grasping, sitting, or opening—has stayed at single-object, single-step interactions. This paper defines a sequential, scene-level version: one instruction such as warming food in a microwave must be decomposed into ordered steps, each grounded as a separate mask over a cluttered 3D Gaussian Splatting (3DGS) scene. It contributes SeqAffordSplat, a benchmark with over 1,800 such scenes and 8,000+ instructions, and SeqSplatNet, an end-to-end model whose language model emits text interleaved with segmentation tokens, each token triggering the decoder to predict one mask. The model reaches 26.2 sIoU in the hardest end-to-end setting versus 12.1 for the point-cloud sequential baseline, and raises the single-step mIoU on the existing 3DAffordSplat benchmark from 30.3 to 40.2.

What carries the argument

The load-bearing mechanism is the interleaved token sequence produced by an autoregressive LLM: text fragments describe each primitive step in words, and each <SEG> token acts as a trigger whose hidden-state vector becomes the query for a lightweight conditional decoder that predicts one binary mask over all Gaussians in the scene. The decoder follows query-based mask prediction, so the language model's per-step hidden state directly conditions the segmentation output. Two supporting mechanisms carry the rest of the argument: Conditional Geometric Reconstruction pretraining, which builds a geometric prior by reconstructing full affordance masks from a mask embedding plus partial geometry, and VFM Semantic Feature Injection, which renders the scene from multiple views, runs those images through a frozen 2D vision foundation model, and lifts the per-pixel features back onto each Gaussian by inverse alpha blending, then fuses them into the decoder at multiple scales.

What would settle it

Test the trained model on real 3DGS scans of furnished rooms with sequential instructions written independently by several humans and annotated without reference to the dataset; if the average overlap score in that setting falls to the 12.1 baseline level, the benchmark's ecological validity claim collapses.

Watch

Extended reading notes

Core claim

The central claim is that sequential affordance reasoning can be cast as autoregressive language generation grounded directly in 3D scenes: a large language model reads the instruction and generates an ordered primitive plan, and every time it emits a special <SEG> token a conditional decoder produces the corresponding 3D affordance mask over the Gaussian primitives. The paper argues that this unifies planning and perception in one end-to-end trainable model, supported by two additions: a pretraining scheme that makes the 3D encoder reconstruct complete affordance regions from geometric observations, and multi-scale injection of semantic features lifted from a frozen 2D vision model through inverse alpha-blended rendering. The evidence is a new large-scale benchmark, SeqAffordSplat, where the method sets the current best results in the single-step setting (37.0 mIoU vs 31.3 for the best prior method), the perception-only sequential setting (36.0 sIoU vs 30.3), and the full end-to-end sequential setting (26.2 sIoU vs 12.1 for the point-cloud SeqAfford baseline). On the existing 3DAffordSplat benchmark it also improves mIoU from 30.3 to 40.2. The paper's conclusion is that high-fidelity 3DGS representations combined with LLM-driven sequential decoding are sufficient to move affordance reasoning from single atomic interactions to long-horizon, scene-level tasks.

Load-bearing premise

The reported gains hold only if scenes composed by combining isolated object models and instructions generated by a language model then edited by humans represent real cluttered environments people and robots encounter.

Editorial extensions

If this is right

  • End-to-end sequential affordance grounding on 3DGS is feasible: the method reaches 26.2 sIoU versus 12.1 for the prior sequential point-cloud method.
  • The unified architecture transfers to single-step tasks, reaching 40.2 mIoU on the 3DAffordSplat benchmark, 9.9 points above the previous published result.
  • Given the ground-truth step sequence, the perception module alone reaches 36.0 sIoU, indicating that geometric pretraining and semantic injection, rather than the LLM's planning, carry the localization accuracy.
  • The ablation shows the geometric pretraining adds roughly 3.8 points of sIoU and DINOv2 feature injection adds about 2.1 more, so both components are needed for the reported gain.
  • LLM scale is not a simple driver of success: the 0.6B backbone outperforms the 8B backbone on sIoU, suggesting that efficient task-tuned language models are sufficient for this type of reasoning.

Reading between the lines

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

  • Because the benchmark scenes are composed by transforming single-object models rather than captured from real rooms, the strongest test of the paper's ecological claim would be a held-out set of real 3DGS room scans with independently written human instructions; the reported gains may not transfer if real clutter and occlusion patterns differ.
  • The <SEG>-token interface suggests a natural bridge to robotics: the same autoregressive planning could emit grasp or placement parameters instead of masks, making the learned sequence directly executable by a manipulator.
  • The observation that the 8B LLM underperforms the 0.6B one hints that the limiting factor is alignment with the curated instruction style rather than world knowledge; a smaller model tuned on a larger pool of human-curated plans may be the more economical path.
  • Because all instructions come from one multimodal language model plus human curation, the diversity of language remains an open question; testing on unseen affordance verbs would separate genuine semantic generalization from memorized phrasings.
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 / 4 minor

Summary. This paper introduces a new task, Sequential 3D Gaussian Affordance Reasoning, together with a benchmark (SeqAffordSplat) of over 1,800 composed 3D Gaussian Splatting scenes annotated with ordered affordance masks and language instructions, and a method (SeqSplatNet) that combines a fine-tuned LLM emitting interleaved text and <SEG> tokens with a conditional 3D affordance decoder, a geometric-reconstruction pre-training, and VFM semantic feature injection. The method is evaluated on the new benchmark (Table 2) and on the existing 3DAffordSplat benchmark (Table 3), reporting sIoU 26.2 vs. 12.1 for SeqAfford in the end-to-end sequential setting and mIoU 40.2 vs. 30.3 on 3DAffordSplat. Ablations (Tables 4-5) quantify the contributions of the pre-training, the VFM feature injection, and the LLM backbone.

Significance. If the empirical claims are supported, the paper makes a useful contribution by extending 3D affordance reasoning from single-object, single-step tasks to multi-object, sequential tasks on 3DGS, and by providing a benchmark that could support future research. The architecture is well motivated, and the ablations demonstrate that the proposed components contribute positively. However, the headline quantitative claims depend on baseline adaptation details and dataset split hygiene that are not currently reported, so the state-of-the-art assertions are not yet established.

major comments (3)
  1. [Experimental Settings (Baseline Models); Table 2] The headline end-to-end Sequential comparison against SeqAfford is not protocol-equivalent. SeqAfford operates on point clouds, and the paper never states how the 3DGS scenes were converted to its input, whether it was retrained on the SeqAffordSplat training split, or which checkpoints and hyperparameters were used. As reported, the 14.1-point sIoU gap (26.2 vs. 12.1) could reflect representation mismatch or an undertrained baseline rather than the superiority of the proposed architecture. The authors should adapt SeqAfford to 3DGS via a clearly specified conversion (e.g., Gaussian centers as point clouds), retrain it under identical data splits, and report the full adaptation and training settings.
  2. [Step 1: 3DGS Data Collection; Table 3] The generalization experiment on 3DAffordSplat is at risk of object-level leakage. SeqAffordSplat scenes are explicitly composed from 3D-AffordanceNet objects, the same corpus underlying 3DAffordSplat, yet the paper does not specify whether the object instances used to build training scenes are disjoint from the 3DAffordSplat test set. If overlapping instances appear, the 9.9-point mIoU improvement (40.2 vs. 30.3) may be inflated by memorization of object geometry rather than by transferable affordance reasoning. Please report the exact object-instance split and demonstrate that no training object from SeqAffordSplat appears in the 3DAffordSplat test split.
  3. [Step 1 and Step 2 of SeqAffordSplat Dataset Collection] The benchmark's ecological validity is asserted but not demonstrated. Scenes are generated by manually composing single-object models with translations, rotations, and scalings, and labels are transferred from single-object annotations; the paper provides no quantitative or qualitative evidence that these compositions reproduce realistic multi-object layouts, occlusion patterns, or object-part interactions, nor does it evaluate the naturalness and executability of the GPT-4o-generated instruction plans beyond reporting curation counts. Because the central claim is state-of-the-art performance on this benchmark, the authors should validate the benchmark, for example by reporting inter-annotator agreement on instruction plans, human ratings of scene realism, or zero-shot transfer to real captured 3DGS scenes.
minor comments (4)
  1. [Abstract / Figure 1 / Dataset section] The scene and annotation counts are inconsistent across the abstract (1800+ scenes, 14,000+ masks, 8,000+ instructions), Figure 1 (over 1,700 scenes, 12,000 instruction pairs), and the Dataset section (1800+ scenes, 14,000+ masks, 8,000+ instructions); please harmonize these numbers.
  2. [Throughout] Typos such as 'envolves', 'acurrate gemoetry', 'Benifiting', 'Obaining', and 'The the lifted semantic features' should be corrected.
  3. [Evaluation Metrics] The sequential metrics align sequences by padding with empty frames; please clarify whether the padding is applied to the predicted or ground-truth sequence and how this interacts with the stated length penalty.
  4. [Tables 2-5] No error bars or repeated-seed results are reported; given that several gains are in the 5-6 point range, please report the variance over at least three seeds.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the reported gains are empirical benchmark results, not by-construction equivalences.

full rationale

The paper's central claims are empirical benchmark outcomes (Tables 2 and 3), not derived predictions. Eq. (1) only defines the task; Eqs. (2)-(7) define losses, a mask-conditioned pretraining objective, and VFM feature lifting, none of which is equivalent to its input by construction. The pretraining stage does take the ground-truth mask M^gt as a conditioning input, but at inference the mask encoder is replaced by the LLM's <SEG> token embedding, so the pretraining is a representation-initialization step rather than a reported prediction. The dataset is assembled from the published 3DAffordSplat/3D-AffordanceNet corpus; although one current co-author (G. Li) is also affiliated with 3DAffordSplat, that corpus is external, published, and independently usable, and the paper also evaluates on the existing 3DAffordSplat benchmark (Table 3), providing an external check. The GPT-4o-generated instruction plans are human-curated, and the method's own LLM is a different model (Qwen3), so no by-construction reduction exists. Concerns about baseline retraining protocol, input conversion, and potential train/test object overlap are experimental-validity/correctness issues, not circularity.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The central claim rests on the validity of the new benchmark, which in turn depends on assumptions about scene composition, label transfer, and the quality of LLM-generated instructions. The method itself introduces no new physical or conceptual entities beyond standard network components. The free parameters are ordinary hyperparameters, but their values are mostly undisclosed, hampering reproduction.

free parameters (2)
  • lambda_mask (loss balancing weight) = not specified
    Hyperparameter in the total loss (Eq. 2) that balances language modeling and mask segmentation losses. Chosen by hand; no value is reported.
  • Number of views m for VFM semantic feature lifting = not specified
    The number of multi-view rendered images used in Eq. (6) and (7) is not stated, yet it directly affects the lifted semantic feature quality and computational cost.
assumptions (5)
  • domain assumption 3DGS representation is sufficient for fine-grained affordance localization.
    The paper assumes that Gaussian primitives with position, rotation, and scale capture the geometry needed for affordance, following 3DAffordSplat. Invoked throughout the method.
  • domain assumption Composing objects from an existing dataset via geometric transformations yields realistic complex scenes.
    In 'SeqAffordSplat Dataset Collection, Step 1', scenes are created by placing objects with translation, rotation, and scaling. This assumes such compositions emulate real-world clutter and multi-object interactions.
  • domain assumption Point-wise affordance labels transfer correctly from source objects to 3DGS Gaussians in composed scenes.
    In 'SeqAffordSplat Dataset Collection, Step 2', labels are projected from point clouds onto the 3DGS representation. This assumes the projection preserves semantic region identity.
  • domain assumption GPT-4o generated instruction sequences, after human curation, are valid ground-truth action plans.
    The benchmark's language supervision comes from MLLM generation with human-in-the-loop curation, as described in Step 2. The validity of these plans as ground truth is assumed.
  • domain assumption Padding predicted and ground-truth sequences to the same length with empty frames is a fair evaluation of sequence quality.
    In the metrics section, sequences are aligned by padding, which penalizes length discrepancies. This assumes that empty frames are an appropriate penalty for both over- and under-generation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SeqAffordSplat: Scene-level Sequential Affordance Reasoning on 3D Gaussian Splatting." pith.science (2026). https://pith.science/paper/YBH7IIME

@misc{pith2026250723772,
  author       = {Pith},
  title        = {Pith review of: SeqAffordSplat: Scene-level Sequential Affordance Reasoning on 3D Gaussian Splatting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YBH7IIME}},
  note         = {Machine review of arXiv:2507.23772}
}
read the original abstract

3D affordance reasoning, the task of associating human instructions with the functional regions of 3D objects, is a critical capability for embodied agents. Current methods based on 3D Gaussian Splatting (3DGS) are fundamentally limited to single-object, single-step interactions, a paradigm that falls short of addressing the long-horizon, multi-object tasks required for complex real-world applications. To bridge this gap, we introduce the novel task of Sequential 3D Gaussian Affordance Reasoning and establish SeqAffordSplat, a large-scale benchmark featuring 1800+ scenes to support research on long-horizon affordance understanding in complex 3DGS environments. We then propose SeqSplatNet, an end-to-end framework that directly maps an instruction to a sequence of 3D affordance masks. SeqSplatNet employs a large language model that autoregressively generates text interleaved with special segmentation tokens, guiding a conditional decoder to produce the corresponding 3D mask. To handle complex scene geometry, we introduce a pre-training strategy, Conditional Geometric Reconstruction, where the model learns to reconstruct complete affordance region masks from known geometric observations, thereby building a robust geometric prior. Furthermore, to resolve semantic ambiguities, we design a feature injection mechanism that lifts rich semantic features from 2D Vision Foundation Models (VFM) and fuses them into the 3D decoder at multiple scales. Extensive experiments demonstrate that our method sets a new state-of-the-art on our challenging benchmark, effectively advancing affordance reasoning from single-step interactions to complex, sequential tasks at the scene level.

Figures

Figures reproduced from arXiv: 2507.23772 by the authors.

Figure 1
Figure 1. (Left)We introduce Sequential 3DGS Affordance Reasoning Task for complex, multi-step agent interactions. (Center)To support this, we present SeqAffordSplat, a large-scale dataset with over 1,700 3DGS scenes and 12,000 instruction pairs.(Right) Our model, SeqSplatNet, sets a new state-of-the-art, improving performance by 6.5% on single-step tasks and 14.1% on our sequential benchmark. Please zoom in for better visual… view at source ↗
Figure 2
Figure 2. An overview of the proposed SeqSplatNet architecture. The architecture comprises four main components: a Large Language Model, a 3DGS Encoder with Conditional Geometric Reconstruction Pre-train, and a Conditional Affordance De￾coder with VFM Semantic Feature Injection. an ordered list of affordance masks, explicitly encoding the temporal and causal order required for the task. To gen￾erate a large and diverse set of… view at source ↗
Figure 3
Figure 3. Visual Results of our proposed methods. ing on 3D Gaussian data. As shown in [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

37 extracted references · 23 canonical work pages

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    L.; Almeida, D.; Altenschmidt, J.; Altman, S.; Anadkat, S.; et al

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

    arXiv 2023

  4. [4]

    Chen, D.; Kong, D.; Li, J.; and Yin, B. 2025. MaskPrompt: Open-Vocabulary Affordance Segmentation with Object Shape Mask Prompts. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, 2034--2042

    work page 2025

  5. [5]

    Cheng, B.; Schwing, A.; and Kirillov, A. 2021. Per-pixel classification is not all you need for semantic segmentation. Advances in neural information processing systems, 34: 17864--17875

    work page 2021

  6. [6]

    Dalgarno, B.; and Lee, M. J. 2010. What are the learning affordances of 3-D virtual environments? British journal of educational technology, 41(1): 10--32

    work page 2010

  7. [7]

    Deng, S.; Xu, X.; Wu, C.; Chen, K.; and Jia, K. 2021. 3d affordancenet: A benchmark for visual object affordance understanding. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 1778--1787

    work page 2021

  8. [8]

    Do, T.-T.; Nguyen, A.; and Reid, I. 2018. Affordancenet: An end-to-end deep learning approach for object affordance detection. In 2018 IEEE international conference on robotics and automation (ICRA), 5882--5889. IEEE

    work page 2018

Show all 37 references
  1. [9]

    J.; Shen, Y.; Wallis, P.; Allen-Zhu, Z.; Li, Y.; Wang, S.; Wang, L.; Chen, W.; et al

    Hu, E. J.; Shen, Y.; Wallis, P.; Allen-Zhu, Z.; Li, Y.; Wang, S.; Wang, L.; Chen, W.; et al. 2022. Lora: Low-rank adaptation of large language models. ICLR, 1(2): 3

    2022

  2. [10]

    Kerbl, B.; Kopanas, G.; Leimk \"u hler, T.; and Drettakis, G. 2023. 3D Gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4): 139--1

    2023

  3. [11]

    P.; and Ba, J

    Kingma, D. P.; and Ba, J. 2014. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980

    2014 arXiv

  4. [12]

    Li, G.; Sun, D.; Sevilla-Lara, L.; and Jampani, V. 2024 a . One-shot open affordance learning with foundation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 3086--3096

    2024

  5. [13]

    Li, Y.; Zhao, N.; Xiao, J.; Feng, C.; Wang, X.; and Chua, T.-s. 2024 b . Laso: Language-guided affordance segmentation on 3d object. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 14251--14260

    2024

  6. [14]

    Lu, D.; Kong, L.; Huang, T.; and Lee, G. H. 2025. Geal: Generalizable 3d affordance learning with cross-modal consistency. In Proceedings of the Computer Vision and Pattern Recognition Conference, 1680--1690

    2025

  7. [15]

    Lu, G.; Zhang, S.; Wang, Z.; Liu, C.; Lu, J.; and Tang, Y. 2024. Manigaussian: Dynamic gaussian splatting for multi-task robotic manipulation. In European Conference on Computer Vision, 349--366. Springer

    2024

  8. [16]

    Marrie, J.; M \'e n \'e gaux, R.; Arbel, M.; Larlus, D.; and Mairal, J. 2024. LUDVIG: Learning-free uplifting of 2d visual features to Gaussian splatting scenes. arXiv preprint arXiv:2410.14462

    2024 arXiv

  9. [17]

    Mo, K.; Qin, Y.; Xiang, F.; Su, H.; and Guibas, L. 2022. O2o-afford: Annotation-free large-scale object-object affordance learning. In Conference on robot learning, 1666--1677. PMLR

    2022

  10. [18]

    S.; Duarte, N

    Mohammadi, S. S.; Duarte, N. F.; Dimou, D.; Wang, Y.; Taiana, M.; Morerio, P.; Dehban, A.; Moreno, P.; Bernardino, A.; Del Bue, A.; et al. 2023. 3dsgrasp: 3d shape-completion for robotic grasp. arXiv preprint arXiv:2301.00866

    2023

  11. [19]

    Nagarajan, T.; and Grauman, K. 2020. Learning affordance landscapes for interaction exploration in 3d environments. Advances in Neural Information Processing Systems, 33: 2005--2015

    2020

  12. [20]

    Oquab, M.; Darcet, T.; Moutakanni, T.; Vo, H.; Szafraniec, M.; Khalidov, V.; Fernandez, P.; Haziza, D.; Massa, F.; El-Nouby, A.; et al. 2023. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193

    2023 arXiv

  13. [21]

    R.; Yi, L.; Su, H.; and Guibas, L

    Qi, C. R.; Yi, L.; Su, H.; and Guibas, L. J. 2017. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30

    2017

  14. [22]

    Qian, S.; Chen, W.; Bai, M.; Zhou, X.; Tu, Z.; and Li, L. E. 2024. Affordancellm: Grounding affordance from vision language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 7587--7597

    2024

  15. [23]

    W.; Hallacy, C.; Ramesh, A.; Goh, G.; Agarwal, S.; Sastry, G.; Askell, A.; Mishkin, P.; Clark, J.; et al

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

    2021

  16. [24]

    Radford, A.; Wu, J.; Child, R.; Luan, D.; Amodei, D.; Sutskever, I.; et al. 2019. Language models are unsupervised multitask learners. OpenAI blog, 1(8): 9

    2019

  17. [25]

    Roy, A.; and Todorovic, S. 2016. A multi-scale cnn for affordance segmentation in rgb images. In European conference on computer vision, 186--201. Springer

    2016

  18. [26]

    Shao, Y.; Zhai, W.; Yang, Y.; Luo, H.; Cao, Y.; and Zha, Z.-J. 2025. Great: Geometry-intention collaborative inference for open-vocabulary 3d object affordance grounding. In Proceedings of the Computer Vision and Pattern Recognition Conference, 17326--17336

    2025

  19. [27]

    Shorinwa, O.; Tucker, J.; Smith, A.; Swann, A.; Chen, T.; Firoozi, R.; Kennedy III, M.; and Schwager, M. 2024. Splat-mover: Multi-stage, open-vocabulary robotic manipulation via editable gaussian splatting. arXiv preprint arXiv:2405.04378

    2024 arXiv

  20. [28]

    H.; Gaskin, J

    Steffen, J. H.; Gaskin, J. E.; Meservy, T. O.; Jenkins, J. L.; and Wolman, I. 2019. Framework of affordances for virtual reality and augmented reality. Journal of management information systems, 36(3): 683--729

    2019

  21. [29]

    C.; Robb, A

    Venkatakrishnan, R.; Venkatakrishnan, R.; Raveendranath, B.; Pagano, C. C.; Robb, A. C.; Lin, W.-C.; and Babu, S. V. 2023. How virtual hand representations affect the perceptions of dynamic affordances in virtual reality. IEEE Transactions on Visualization and Computer Graphic...

    2023

  22. [30]

    Wei, Z.; Lin, J.; Liu, Y.; Chen, W.; Luo, J.; Li, G.; and Lin, L. 2025. 3DAffordSplat: Efficient Affordance Reasoning with 3D Gaussians. arXiv preprint arXiv:2504.11218

    2025 arXiv

  23. [31]

    Xu, C.; Chen, Y.; Wang, H.; Zhu, S.-C.; Zhu, Y.; and Huang, S. 2022. Partafford: Part-level affordance discovery from 3d objects. arXiv preprint arXiv:2202.13519

    2022 arXiv

  24. [32]

    G.; Petit, D.; Tsuji, T.; Akizuki, S.; Hashimoto, M.; Nagata, K.; and Harada, K

    Yamanobe, N.; Wan, W.; Ramirez-Alpizar, I. G.; Petit, D.; Tsuji, T.; Akizuki, S.; Hashimoto, M.; Nagata, K.; and Harada, K. 2017. A brief review of affordance in robotic manipulation research. Advanced Robotics, 31(19-20): 1086--1101

    2017

  25. [33]

    Yang, A.; Li, A.; Yang, B.; Zhang, B.; Hui, B.; Zheng, B.; Yu, B.; Gao, C.; Huang, C.; Lv, C.; et al. 2025. Qwen3 technical report. arXiv preprint arXiv:2505.09388

    2025 arXiv

  26. [34]

    Yang, Y.; Zhai, W.; Luo, H.; Cao, Y.; Luo, J.; and Zha, Z.-J. 2023. Grounding 3d object affordance from 2d interactions in images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 10905--10915

    2023

  27. [35]

    Yu, C.; Wang, H.; Shi, Y.; Luo, H.; Yang, S.; Yu, J.; and Wang, J. 2025. Seqafford: Sequential 3d affordance reasoning via multimodal large language model. In Proceedings of the Computer Vision and Pattern Recognition Conference, 1691--1701

    2025

  28. [36]

    Zheng, Y.; Chen, X.; Zheng, Y.; Gu, S.; Yang, R.; Jin, B.; Li, P.; Zhong, C.; Wang, Z.; Liu, L.; et al. 2024. Gaussiangrasper: 3d language gaussian splatting for open-vocabulary robotic grasping. IEEE Robotics and Automation Letters

    2024

  29. [37]

    Zhu, S.; Wang, G.; Kong, X.; Kong, D.; and Wang, H. 2024. 3d gaussian splatting in robotics: A survey. arXiv preprint arXiv:2410.12262

    2024 arXiv

Pith tools

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