Pith. sign in

REVIEW 4 major objections 5 minor 37 references

Speech2Grasp: Data-Efficient Transfer of Text-Conditioned Grasp Detection to Speech in Humanoid Robots

T0 review · 4 major / 5 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read A text-conditioned grasp detector can be transferred to speech by training a small MLP projector on about 15K image-speech pairs, letting a humanoid grasp objects directly from spoken commands without an ASR step, with equal or better succe

desk verdict Useful transfer result, real-robot gap is meaningful, but the paper's own evidence for instruction-level fidelity is thinner than it looks. read the letter →

arxiv 2607.26567 v1 pith:RNQKFATA submitted 2026-07-29 cs.RO cs.CV

Hung Nguyen , Kim Nhat Minh Nguyen , Van Duc Vu , Van-Danh Le , Hoang Huy Le , Dinh Tuan Nguyen , Pham Tuyen Le , Van-Truong Nguyen
show 1 more author
Quan Nguyen
This is my paper
classification cs.ROcs.CV
keywords speech-conditionedgraspingmodalitytransferknowledgedistillationvision-languagemodelshumanoidroboticsspeechrepresentationalignmentgraspdetection
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 asks whether an established text-conditioned vision-language model can be made to accept speech directly, without large speech datasets or architectural redesign. It answers yes: with ALBEF as the test bed, a lightweight MLP projector trained with a cosine-similarity knowledge-distillation loss maps Whisper speech embeddings into the ALBEF fused embedding space, lifting pair-matched similarity to near-perfect alignment while preserving discrimination between matched and random speech. The resulting speech-conditioned grasp detector, Speech2Grasp, matches the text teacher in simulation and beats a cascaded ASR-based pipeline on a real humanoid robot, because it never transcribes and therefore never suffers discrete mistranscription errors. The win matters because it suggests a general recipe for extending text-conditioned robotic perception to the more natural speech modality at small data and compute cost.

What carries the argument

The load-bearing mechanism is the projected [CLS] embedding: the frozen ALBEF fusion produces a single global [CLS] vector representing the image-text joint semantics, and the student branch is trained so that the image-speech [CLS] vector matches it under cosine similarity (Eq. 1). A lightweight MLP with a skip connection maps Whisper-pooled speech tokens into this space; the same projected pseudo-text tokens are then injected into ALBEF's fusion bottleneck, letting the frozen text-conditioned grasp head act on speech. An off-ramp scheduler shifts from the distillation loss to the grasp task loss, and a DWT-based frequency-aware encoder adds camera-noise robustness.

What would settle it

Fix an image of an object with two distinctive graspable parts and generate two spoken instructions that differ only in the part name (e.g., 'by its bowl shape' vs 'by its handle'). If the model produces nearly identical grasp maps for the two, or if per-instruction success on such minimal pairs is near chance while aggregate success remains high, then the CLS-level alignment is insufficient and the transfer claim fails.

Watch

Extended reading notes

Core claim

The paper claims that a frozen text-conditioned grasp detector can be transferred to speech by training only a lightweight MLP projector that maps Whisper speech embeddings into ALBEF's fused [CLS] embedding space, using a cosine-similarity knowledge-distillation loss on about 15K image-text-speech triples. The projection raises pair-matched cosine similarity from 0.35 to 0.97 while keeping randomly paired speech at 0.84, preserving semantic discrimination. In simulation the speech-conditioned model matches the text teacher (harmonic success 0.35 vs 0.35) and in real-world humanoid experiments it outperforms a cascaded ASR pipeline (single-object 0.70 vs 0.59, multi-object 0.61 vs 0.54) whil

Load-bearing premise

The whole transfer rests on the assumption that ALBEF's single [CLS] embedding—a global summary of the image-text pair—carries enough fine-grained semantic detail to drive grasp selection, so aligning that summary for speech is sufficient.

Editorial extensions

If this is right

  • Speech2Grasp matches or exceeds the text-conditioned teacher in simulation (harmonic success 0.35 vs 0.35) while using only 15K speech samples rather than the million text samples used for the teacher.
  • Because the ASR step is removed, inference latency drops from about 102 ms to about 37 ms for producing the conditioning embedding.
  • The same recipe could port other text-conditioned perception models (segmentation, detection, pick-and-place) to speech with minimal new data, as long as they expose a pooled joint embedding.
  • The projected embeddings remain aligned under unseen speakers and acoustic corruption (worst-case mean cosine similarity 0.915), suggesting deployability in noisy offices.
  • The DWT visual branch maintains grasp success under Poisson-Gaussian camera noise, where a non-DWT pipelined counterpart degrades.

Reading between the lines

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

  • The alignment target is the global [CLS] vector, so the transfer may be insensitive to instruction details that are lost in whole-utterance pooling; a minimal-pair test (bowl vs handle for the same object) would reveal whether the projector preserves the fine-grained distinctions grasp detection needs.
  • If the recipe generalizes, it suggests that modality transfer can be decoupled from task learning: any frozen text-conditioned robot skill could be exposed to speech by matching pooled embeddings rather than retraining on paired robotic speech data.
  • The paper's latency comparison uses a lightweight ASR; a stronger ASR would correct more transcriptions but would likely increase latency further, so the reported advantage may be conservative on accuracy and optimistic on latency depending on the ASR.
  • A natural extension is to test whether the projector learned a genuinely continuous mapping—so that novel paraphrases or code-switched speech land near their text counterparts—or merely memorized a fixed vocabulary from the 15K prompts.
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 addresses the problem of extending a text-conditioned grasp-detection model (ALBEF-based LGD) to accept speech inputs directly, without an ASR transcription step. The authors first perform diagnostic analyses showing that a lightweight MLP projector trained with a cosine-similarity distillation loss (Eq. 1) on the fused [CLS] embeddings of image–speech and image–text pairs can closely align speech-conditioned representations with text-conditioned ones. They then introduce Speech2Grasp, which combines this projector with a frozen/lightly fine-tuned grasp head and a DWT-based vision encoder. Experiments in simulation on Grasp-Anything++ and on a real humanoid robot compare Speech2Grasp with a LiteASR→LGD cascade, reporting comparable or better grasp success rates and roughly threefold lower latency for producing the ALBEF [CLS] embedding, while requiring only 15K speech samples.

Significance. If the central claim is established, the paper offers a practical and data-efficient recipe for retrofitting established text-conditioned vision-language robotic systems with speech input, an important step for natural human–robot interaction. The real-robot experiments are a genuine strength, as is the explicit use of a small-scale adaptation set and the ablation on training sample count (Table III). However, the load-bearing evidence for instruction-following capability is currently aggregate success rates and a cosine-similarity metric that may not reflect fine-grained semantic distinctions. The paper's significance therefore depends on closing this gap with per-instruction evaluation and statistical rigor.

major comments (4)
  1. [Section III-B and Section V-A, Eq. (1), Tables I–III] The claim that Speech2Grasp preserves the teacher's grasp capability despite bypassing transcription is not yet supported. The diagnostic evidence for semantic discrimination (Fig. 3a) compares matched image–speech–text triples against randomly paired speech from different samples, so both the image and the instruction differ; a model that encodes only the image/sample identity could produce the reported gap. More importantly, the downstream evaluation reports only aggregate success rates on S/U categories. A model that ignores the instruction and produces an object-appropriate generic grasp could attain similar aggregate success, particularly given the teacher's low harmonic mean (H=0.35) and the short, simple prompts. The paper should report per-instruction correctness, e.g., success conditioned on contrasting prompts for the same object ("by its bowl shape" vs "by its handle"), or at
  2. [Section V-A, Tables I–III] Tables I–III report results 'averaged across 10 re-runs' but provide no error bars, confidence intervals, or significance tests. The simulation advantage over the ASR cascade is H=0.35 vs H=0.34, which may be within run-to-run variability. The real-robot results (Table V) do include standard errors, but with 30 attempts per condition per speaker, a significance test (e.g., a two-proportion test or bootstrap) is needed to support the headline claim of outperformance. Please add error bars to all re-run tables and report statistical significance for the main comparisons.
  3. [Section V-A, Table I] The latency metric is described as the time 'to produce the ALBEF [CLS] embedding' for the same speech. This is not an end-to-end comparison of the two pipelines. The full pipeline includes image encoding, fusion, decoding, and grasp-head execution; for the ASR baseline, the ASR transcription time must also be counted. Reporting only the embedding generation time overstates the latency advantage. Please report end-to-end inference latency for both systems.
  4. [Section III-B] The train-validation-test split is stated as '80%-20%-20%', which sums to 120%. The subsequent text in Section III-D refers to '9K paired image-speech training samples', which implies a 60/20/20 split (9K of 15K). Please correct this inconsistency and ensure the data-efficiency claims (15K samples, 9K training) are based on a coherent split.
minor comments (5)
  1. [Section V-A, Table I] The column header '#S(↓)' is used for both the number of training samples (15K for Speech2Grasp) and 1M for LiteASR→LGD and LGD. For the ASR baseline, 1M likely refers to ASR training data, while for LGD it likely refers to text-grasp training data; please define what '#S' means for each row.
  2. [Section III-C] The text says the projector is 'trained exclusively on clean speeches of a single voice' but then evaluates on 'seven unseen voices'. Please clarify the number of training voices and the identity of the 'Jasper' voice relative to the seven unseen ones.
  3. [Section IV-C] The 'off-ramp linear scheduler' for λ_KD is not defined. Please specify the schedule, including initial and final values and the number of epochs/iterations over which it is annealed.
  4. [Section V-A, Table II] The DWT contribution is evaluated only by comparing Speech2Grasp (with DWT) against baselines without DWT. To isolate the effect of discarding the HH subband, please include an ablation of Speech2Grasp without DWT (or with all four subbands) under the same noise conditions.
  5. [General] The real-robot experiments use 'two non-native English speakers' but Table V reports results per person. Please state whether the speakers are the same across single-object and multi-object conditions, and whether the 30 repetitions per case are independent or repeated on the same objects (which affects the validity of the reported standard errors).

Circularity Check

1 steps flagged · score 4.0 of 10

Partial circularity: the diagnostic 'adaptation' evidence is the KD loss itself; grasp results are independent.

  1. fitted input called prediction [Section III-B, Eq. (1) and the paragraph reporting CS = 0.968]
    "The projector maps the speech embeddings into the shared ALBEF latent space, and is trained through the following objective based on cosine similarity (CS): LKD = 1− e⊤S eT / ∥eS∥2 ∥eT∥2 (1) ... The blue curve indicates that the trained projector successfully maps the speech embeddings into the teacher’s latent space. The average CS is 0.968±0.027, indicating a very strong alignment using only a lightweight MLP-based projector."

    The projector is trained by minimizing 1 minus the cosine similarity between student and teacher [CLS] embeddings (Eq. 1). The diagnostic then reports the resulting cosine similarity (0.968) as evidence that the projector 'effectively adapts' ALBEF to speech. This metric is exactly the training objective; reporting it on matched pairs measures how well the model fit the loss, not an independent confirmation of adaptation. The downstream grasp success provides independent support, but the paper's 'central finding' about the projector's sufficiency rests partly on this by-construction alignment.

full rationale

The only clear circularity is in the diagnostic study: the projector is trained with the cosine-similarity knowledge-distillation loss (Eq. 1), and the same cosine similarity is then presented as the primary evidence that the projector 'effectively adapts' ALBEF to speech. This is a fitted-input-called-prediction pattern, though it is not the paper's final claim. The final claim that Speech2Grasp outperforms a cascaded ASR pipeline is supported by simulation and real-robot grasp success rates, which are external to the KD objective and thus non-circular. The DWT component is introduced via several self-citations [26]-[30], but its benefit is empirically evaluated in Table II (with/without DWT), so that self-citation is not load-bearing. No uniqueness theorem or ansatz-smuggling is present. The semantic-discrimination comparison using random pairs is not circular (the model was not trained on negative pairs), though it is confounded by the image also differing between matched and random pairs; that is a correctness risk rather than a circularity. Overall, the central derivation is largely self-contained, with one partial circularity in the diagnostic motivation, warranting a score of 4.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The paper does not postulate new physical or conceptual entities; the 'speech projector' and 'pseudo-text token sequence' are trained architectural components, not independently hypothesized entities. The main unstated costs are the λ_KD schedule and data budget, plus domain assumptions about Whisper, ALBEF's CLS bottleneck, synthetic speech, and DWT suitability.

free parameters (3)
  • λ_KD balancing term = Not reported
    In Eq. 3, λ_KD weights the knowledge distillation loss against the grasp loss, with an off-ramp linear scheduler. No values or schedule endpoints are given, and no sensitivity analysis is provided. The choice affects how strongly speech embeddings are pushed toward text embeddings and therefore the final grasp performance.
  • Training sample count (15K) = 15K
    Table III shows the system is evaluated at 7.5K, 11K, and 15K samples; 15K gives the best results and is used for the main experiments. This is a data-budget selection made after seeing performance, not a principled criterion.
  • DWT subband selection = LL, LH, HL kept; HH discarded
    The frequency-aware encoder hand-selects which wavelet subbands to keep, discarding HH as noise-dominant. This design choice is motivated by prior work and is not fitted, but it is a free architectural assumption that affects robustness results.
assumptions (5)
  • domain assumption Whisper speech embeddings capture enough semantic content to be alignable with text embeddings via a simple projector.
    The entire diagnostic study (Section III-B/C) and framework (Section IV-B) rely on Whisper's pre-trained representations containing the semantic information needed for grasp instructions. If Whisper embeddings lose instruction details, no projector can recover them.
  • domain assumption The ALBEF [CLS] fused embedding is a sufficient bottleneck for language-driven grasp detection, so CLS-level alignment transfers grasp capability.
    Eq. 1 trains only the CLS-level alignment, and the framework injects the projected speech CLS into the 'semantic bottleneck' for vision-language fusion (Section IV-B). This assumes fine-grained token-level information is not required for correct grasping.
  • domain assumption Synthetic TTS speech from Kitten-TTS and the Grasp-Anything++ prompts is representative of natural user commands and acoustic conditions.
    Training uses Kitten-TTS synthesized speech with a single voice (Jasper). Real-world evaluation later uses two non-native English speakers, but the training distribution is synthetic and may not cover the full command and pronunciation space.
  • domain assumption The LGD teacher's fusion module can accept a 'pseudo-text' sequence of Whisper-derived tokens in place of BERT tokens.
    Section IV-C and Figure 5 assume that replacing the text token sequence with pooled Whisper output tokens ([SUMMARY] plus [UTTER]) is structurally compatible with ALBEF's fusion, without architectural changes.
  • domain assumption Discarding the HH wavelet subband removes camera noise without losing object-boundary information needed for grasping.
    Section IV-D justifies this via prior wavelet works [26]-[31], but the paper does not provide a controlled ablation within the same architecture to verify the DWT choice independently of the speech-transfer claims.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Speech2Grasp: Data-Efficient Transfer of Text-Conditioned Grasp Detection to Speech in Humanoid Robots." pith.science (2026). https://pith.science/paper/RNQKFATA

@misc{pith2026260726567,
  author       = {Pith},
  title        = {Pith review of: Speech2Grasp: Data-Efficient Transfer of Text-Conditioned Grasp Detection to Speech in Humanoid Robots},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RNQKFATA}},
  note         = {Machine review of arXiv:2607.26567}
}
read the original abstract

Humanoid robots increasingly require multi-modal understanding for natural interaction with humans. Despite the prominence of vision-language models, they generally assume textual rather than the more natural speech inputs. In this paper, we investigate whether a well-established text-conditioned model can be transferred to speech in a data-efficient manner. Using ALBEF as a case study, we conduct diagnostic analyses showing that a lightweight MLP-based projector effectively adapts it to speech, while preserving semantic discrimination and robustness. Motivated by these findings, we introduce Speech2Grasp, a framework for data-efficient transfer of text-conditioned grasp detection to speech. Real-world humanoid robot experiments show that Speech2Grasp outperforms cascaded ASR-based pipeline, while reducing inference latency. Our findings suggest a practical paradigm for extending established text-conditioned systems to speech.

Figures

Figures reproduced from arXiv: 2607.26567 by the authors.

Figure 1
Figure 1. Methods for transferring text-conditioned models to speech. a) uses [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. ALBEF [9] fusion between image, text and speech modalities. The [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Effects of the trained projector on text-speech representation gap. [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (4 more)
Figure 5
Figure 5. Figure 5: Overview of the Speech2Grasp framework for data-efficient transfer of text-conditioned grasp detection to speech. An MLP-based projector [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: Speech encoding and projection. The augmented speech is converted [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 7
Figure 7. Figure 7: The DWT decomposes a noisy input image into four wavelet [PITH_FULL_IMAGE:figures/full_fig_p006_7.png]
Figure 8
Figure 8. Figure 8: Grasp sequence of the humanoid robot using Speech2Grasp. [PITH_FULL_IMAGE:figures/full_fig_p007_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

37 extracted references · 4 linked inside Pith

  1. [1]

    A survey of multimodal perception methods for human–robot interaction in social environments,

    J. A. Duncan, F. Alambeigi, and M. W. Pryor, “A survey of multimodal perception methods for human–robot interaction in social environments,”J. Hum.-Robot Interact., vol. 13, no. 4, Oct. 2024. [Online]. Available: https://doi.org/10.1145/3657030

    doi:10.1145/3657030 2024

  2. [2]

    Bridging language and action: A survey of language-conditioned robot manipulation,

    X. Yao, H. Zhou, O. Mees, Y . Meng, T. Xiao, Y . Bisk, J. Oh, E. Johns, M. Shridhar, D. Shah, J. Thomason, K. Huang, J. Chai, Z. Bing, and A. Knoll, “Bridging language and action: A survey of language-conditioned robot manipulation,” 2026. [Online]. Available: https://arxiv.org/abs/2312.10807

    arXiv 2026

  3. [3]

    Imagebind: One embedding space to bind them all,

    R. Girdhar, A. El-Nouby, Z. Liu, M. Singh, K. V . Alwala, A. Joulin, and I. Misra, “Imagebind: One embedding space to bind them all,” in CVPR, 2023

    2023

  4. [4]

    Languagebind: Extending video-language pretraining to n-modality by language-based semantic alignment,

    B. Zhu, B. Lin, M. Ning, Y . Yan, J. Cui, W. HongFa, Y . Pang, W. Jiang, J. Zhang, Z. Li, C. W. Zhang, Z. Li, W. Liu, and L. Yuan, “Languagebind: Extending video-language pretraining to n-modality by language-based semantic alignment,” 2023

    2023

  5. [5]

    Onellm: One framework to align all modalities with language,

    J. Han, K. Gong, Y . Zhang, J. Wang, K. Zhang, D. Lin, Y . Qiao, P. Gao, and X. Yue, “Onellm: One framework to align all modalities with language,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024

    2024

  6. [6]

    Large-scale contrastive language-audio pretraining with feature fusion and keyword-to-caption augmentation,

    Y . Wu*, K. Chen*, T. Zhang*, Y . Hui*, T. Berg-Kirkpatrick, and S. Dubnov, “Large-scale contrastive language-audio pretraining with feature fusion and keyword-to-caption augmentation,” inIEEE In- ternational Conference on Acoustics, Speech and Signal Processing, ICASSP, 2023

    2023

  7. [7]

    SSR: Alignment-aware modality connector for speech language models,

    W. Tan, H. Inaguma, N. Dong, P. D. Tomasello, and X. Ma, “SSR: Alignment-aware modality connector for speech language models,” inProceedings of the 22nd International Conference on Spoken Language Translation (IWSLT 2025), E. Salesky, M. Federico, and A. Anastasopoulos, Eds. Vienna, Austria (in-person and online): Association for Computational Linguistics...

    2025

  8. [8]

    End-to-end speech recognition: A survey,

    R. Prabhavalkar, T. Hori, T. N. Sainath, R. Schl ¨uter, and S. Watanabe, “End-to-end speech recognition: A survey,”IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 32, pp. 325–351, 2024

    2024

Show all 37 references
  1. [9]

    Align before fuse: vision and language representation learning with momentum distillation,

    J. Li, R. R. Selvaraju, A. D. Gotmare, S. Joty, C. Xiong, and S. C. Hoi, “Align before fuse: vision and language representation learning with momentum distillation,” inProceedings of the 35th International Conference on Neural Information Processing Systems, ser. NIPS ’21. Red...

    2021

  2. [10]

    Grasp-anything: Large-scale grasp dataset from foundation models,

    A. D. Vuong, M. N. Vu, H. Le, B. Huang, H. T. T. Binh, T. V o, A. Kugi, and A. Nguyen, “Grasp-anything: Large-scale grasp dataset from foundation models,” in2024 IEEE International Conference on Robotics and Automation (ICRA), 2024, pp. 14 030–14 037

    2024

  3. [11]

    Adaptive knowledge distillation between text and speech pre-trained models,

    J. Ni, Y . Ma, W. Wang, Q. Chen, D. Ng, H. Lei, T. H. Nguyen, C. Zhang, B. Ma, and E. Cambria, “Adaptive knowledge distillation between text and speech pre-trained models,”ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp....

    2023

  4. [12]

    Afd-slu: Adaptive feature distillation for spoken language understanding,

    Y . Xie, Y . Cui, L. Xie, and E. Yin, “Afd-slu: Adaptive feature distillation for spoken language understanding,” inICASSP 2026 - 2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2026, pp. 19 432–19 436

    2026

  5. [13]

    Smaller and faster robotic grasp detection model via knowledge distillation and unequal feature encoding,

    H. Nie, Z. Zhao, L. Chen, Z. Lu, Z. Li, and J. Yang, “Smaller and faster robotic grasp detection model via knowledge distillation and unequal feature encoding,”IEEE Robotics and Automation Letters, vol. 9, no. 8, pp. 7206–7213, 2024

    2024

  6. [14]

    Lite- grasp: A light robotic grasp detection via semi-supervised knowledge distillation,

    L. Peng, R. Cai, J. Xiang, J. Zhu, W. Liu, W. Gao, and Y . Liu, “Lite- grasp: A light robotic grasp detection via semi-supervised knowledge distillation,”IEEE Robotics and Automation Letters, vol. 9, no. 9, pp. 7995–8002, 2024

    2024

  7. [15]

    Pdcnet: A lightweight and efficient robotic grasp detection framework via partial convolution and knowledge distillation,

    Y . Jiang, Y . Fang, and L. Deng, “Pdcnet: A lightweight and efficient robotic grasp detection framework via partial convolution and knowledge distillation,”Computer Vision and Image Understanding, vol. 259, p. 104441, 2025. [Online]. Available: https://www.sciencedirect.com/s...

    2025

  8. [16]

    Language-driven grasp detection,

    A. D. Vuong, M. N. Vu, B. Huang, N. Nguyen, H. Le, T. V o, and A. Nguyen, “Language-driven grasp detection,” in2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024, pp. 17 902–17 912

    2024

  9. [17]

    Graspsam: When segment anything model meets grasp detection,

    S. Noh, J. Kim, D. Nam, S. Back, R. Kang, and K. Lee, “Graspsam: When segment anything model meets grasp detection,” in2025 IEEE International Conference on Robotics and Automation (ICRA), 2025, pp. 14 023–14 029

    2025

  10. [18]

    Segment anything,

    A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. Whitehead, A. C. Berg, W.-Y . Lo, P. Doll ´ar, and R. Gir- shick, “Segment anything,”arXiv:2304.02643, 2023

    2023 arXiv

  11. [19]

    Language-driven 6-dof grasp detection using negative prompt guidance,

    T. Nguyen, M. N. Vu, B. Huang, A. Vuong, Q. Vuong, N. Le, T. V o, and A. Nguyen, “Language-driven 6-dof grasp detection using negative prompt guidance,” inECCV, 2024

    2024

  12. [20]

    Kittentts: State-of-the-art lightweight text-to-speech model,

    KittenML, “Kittentts: State-of-the-art lightweight text-to-speech model,” GitHub repository, 2026, https://github.com/KittenML/KittenTTS

    2026

  13. [21]

    Robust speech recognition via large-scale weak supervi- sion,

    A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, “Robust speech recognition via large-scale weak supervi- sion,” inInternational Conference on Machine Learning, 2022. [On- line]. Available: https://api.semanticscholar.org/CorpusID:252923993

    2022

  14. [22]

    Visualizing data using t-SNE,

    L. van der Maaten and G. Hinton, “Visualizing data using t-SNE,”Journal of Machine Learning Research, vol. 9, pp. 2579–2605, 2008. [Online]. Available: http://www.jmlr.org/papers/v9/vandermaaten08a.html

    2008

  15. [23]

    A study on data augmentation of reverberant speech for robust speech recognition,

    T. Ko, V . Peddinti, D. Povey, and S. Khudanpur, “A study on data augmentation of reverberant speech for robust speech recognition,” inIEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2017

    2017

  16. [24]

    Jurafsky and J

    D. Jurafsky and J. H. Martin,Speech and Language Processing, 2nd ed. Prentice Hall, 2009

    2009

  17. [25]

    Bert: Pre-training of deep bidirectional transformers for language understanding,

    J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “Bert: Pre-training of deep bidirectional transformers for language understanding,” inNorth American Chapter of the Association for Computational Linguistics, 2019. [Online]. Available: https://api.semanticscholar.org/CorpusID:52967399

    2019

  18. [26]

    Dwtnerf: Boosting few-shot neural radiance fields via discrete wavelet transform,

    H. Nguyen, B. R. Li, and T. Nguyen, “Dwtnerf: Boosting few-shot neural radiance fields via discrete wavelet transform,” 2025. [Online]. Available: https://arxiv.org/abs/2501.12637

    2025 arXiv

  19. [27]

    Dwtgs: Rethinking frequency regularization for sparse-view 3d gaussian splatting,

    H. Nguyen, R. Li, A. Le, and T. Nguyen, “Dwtgs: Rethinking frequency regularization for sparse-view 3d gaussian splatting,” in Proceedings of the 2025 IEEE International Conference on Visual Communications and Image Processing (VCIP). IEEE, 2025

    2025

  20. [28]

    From coarse to fine: Learnable discrete wavelet transforms for efficient 3d gaussian splatting,

    H. Nguyen, A. Le, B. R. Li, and T. Nguyen, “From coarse to fine: Learnable discrete wavelet transforms for efficient 3d gaussian splatting,” inProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, October 2025, pp. 3139–3148

    2025

  21. [29]

    Waveletgaussian: Wavelet- domain diffusion for sparse-view 3d gaussian object reconstruction,

    H. Nguyen, R. Li, A. Le, and T. Nguyen, “Waveletgaussian: Wavelet- domain diffusion for sparse-view 3d gaussian object reconstruction,” in Proceedings of the 2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2026

    2026

  22. [30]

    Learnable multi-level discrete wavelet transforms for 3d gaussian splatting frequency modulation,

    H. Nguyen, A. Le, and T. Nguyen, “Learnable multi-level discrete wavelet transforms for 3d gaussian splatting frequency modulation,”

  23. [31]

    Wavelet diffusion models are fast and scalable image generators,

    H. Phung, Q. Dao, and A. Tran, “Wavelet diffusion models are fast and scalable image generators,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2023, pp. 10 199–10 208

    2023

  24. [32]

    Lvis: A dataset for large vocabulary instance segmentation,

    A. Gupta, P. Doll ´ar, and R. Girshick, “Lvis: A dataset for large vocabulary instance segmentation,” in2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019, pp. 5351– 5359

    2019

  25. [33]

    Conditional prompt learning for vision-language models,

    K. Zhou, J. Yang, C. C. Loy, and Z. Liu, “Conditional prompt learning for vision-language models,” in2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022, pp. 16 795–16 804

    2022

  26. [34]

    Liteasr: Efficient automatic speech recognition with low-rank approximation,

    K. Kamahori, J. Kasai, N. Kojima, and B. Kasikci, “Liteasr: Efficient automatic speech recognition with low-rank approximation,” inPro- ceedings of the 2025 Conference on Empirical Methods in Natural Language Processing (EMNLP). Association for Computational Linguistics, 2025,...

    2025

  27. [35]

    Pogain: Poisson-gaussian image noise modeling from paired sam- ples,

    N. B ¨ahler, M. El Helou, ´E. Objois, K. Okumus ¸, and S. S ¨usstrunk, “Pogain: Poisson-gaussian image noise modeling from paired sam- ples,”IEEE Signal Processing Letters, vol. 29, pp. 2602–2606, 2022

    2022

  28. [36]

    Unitree robotics official website,

    Unitree Robotics, “Unitree robotics official website,” https://www.unitree.com/, n.d., accessed: 2026-03-03

    2026

  29. [2026]

    Available: https://arxiv.org/abs/2602.14199

    [Online]. Available: https://arxiv.org/abs/2602.14199

    arXiv

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.