Pith. sign in

REVIEW 4 major objections 3 minor 34 references

A humanoid robot can listen to music or speech and autonomously choose which whole-body motion skill to perform, with 84.8% chunk-level accuracy in simulation and a successful transfer to the physical robot.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-02 03:25 UTC pith:N4WOKVKF

load-bearing objection A credible integration demo whose 'semantic' claim is not actually tested; the 84.8% measures fingerprint lookup on four known songs. the 4 major comments →

arxiv 2607.14182 v1 pith:N4WOKVKF submitted 2026-07-15 cs.RO cs.AI

Semantic Audio-driven Understanding for Dynamic Humanoid Whole Body Control

J. M. A. Marcelo , M. Brienza , E. Bugli , L. Comito , D. Nardi , D. D. Bloisi , V. Suriani This is my paper
classification cs.RO cs.AI
keywords humanoid robotsaudio-driven controlwhole-body controlsemantic groundingmotion skillsretrievalsim-to-real transferspeech interaction
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tries to show that a humanoid robot can listen to a continuous audio stream—music or speech—and autonomously decide which whole-body motion skill to execute, without pre-scripted timing or human-triggered cues. It reports 84.8% chunk-level accuracy in simulation when the retrieved policy matches the authors' hand-authored timed skill rules, and it demonstrates the same control logic on a physical humanoid. If correct, this replaces fixed choreography with perception-driven behavior selection, moving robots from 'perform this script' to 'respond to what is happening around them.' The system's core move is to split audio into fixed chunks, route each chunk to a music or speech branch, and map the retrieved identity or transcript to a discrete skill library via a unified scheduler.

Core claim

The central claim is that semantic audio understanding—not just audio-triggered cues—can drive whole-body humanoid control. Music chunks are identified by fingerprinting and semantic embeddings; the estimated time offset within a track selects a skill through timed rules; speech chunks are transcribed and matched against a skill library. Both branches dispatch skill identifiers to a shared policy-scheduling layer. The authors report 84.8% overall chunk-level accuracy over 574 evaluated chunks under shifted boundary conditions, and qualitatively confirm on the physical robot that the commanded policy sequence is followed.

What carries the argument

The load-bearing component is the retrieval-and-routing chain: a hierarchical router separates music from speech, a fingerprint module with an embedding fallback identifies the track and its temporal offset, and a resolver maps (track, offset) or transcript to a skill identifier via timed skill rules, which are structured as a reference track, a time interval, and a target policy. The unified interface then schedules these identifiers over imitation-learned whole-body policies, with a standing policy mediating transitions. This lets different sections of a song—intro, verse, chorus, outro—drive different behaviors without retraining the underlying controllers.

Load-bearing premise

The 84.8% accuracy is measured against the authors' own hand-authored timed skill rules that assign each audio segment to one policy, so the number shows how consistently the system matches those rules, not whether the mappings are objectively right or generalize to other songs.

What would settle it

Run the system on a set of songs not in its reference catalogue, with ground truth established by human raters who score whether each selected motion fits the audio; if retrieval accuracy or human agreement collapses, the 'semantic understanding' claim fails even though the fingerprint-based lookup still works for known tracks.

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

If this is right

  • Robots can perform live choreography that adapts to which song is playing and where in the song it is, rather than replaying a fixed timeline.
  • A human can command the robot through speech using the same pipeline, enabling direct interaction and mixed-initiative control.
  • Because skill selection happens at retrieval time through embeddings, adding a new motion skill does not require retraining existing policies; it only extends the library and its descriptors.
  • The same scheduling interface works in simulation and on real hardware unchanged, so simulation results can predict real-world behavior with reasonable fidelity.
  • If the reported accuracy generalizes beyond the four tested tracks, the approach applies to any performance setting where audio context should drive behavior.

Where Pith is reading between the lines

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

  • The 84.8% figure is a consistency measure against the authors' own hand-written timed rules; a fair test of 'semantic' appropriateness would ask independent viewers whether the chosen policies match the audio.
  • The fixed 5-second chunks and the stand-primed transition create a lower bound on reaction latency; an adaptive segmentation module, which the authors list as future work, could make the robot respond faster to musical changes.
  • The same routing-and-retrieval pattern could be applied to other continuous modalities—video, motion capture, or sensor streams—since the core is mapping chunks to a discrete skill library rather than anything audio-specific.
  • Fingerprinting handles known tracks exactly, while the embedding fallback is meant to handle novel audio; measuring performance on never-seen songs would reveal how much of the 'understanding' is semantic versus memorized identity.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 3 minor

Summary. The paper proposes a multi-modal orchestration framework that maps continuous audio to whole-body motion skills for a Unitree G1 humanoid. The system routes each 5 s audio chunk into music or speech branches: music is matched via Wang audio fingerprinting with a CLAP embedding fallback, and speech is transcribed and grounded into a skill library; a resolver then selects a policy based on hand-authored timed skill rules and schedules execution over imitation-learned RL policies. The authors report 84.8% chunk-level retrieval accuracy over 574 shifted chunks from four songs, demonstrate simulated mashup execution at 20 s and 30 s granularity, and provide a qualitative real-robot deployment.

Significance. If the claims are fully supported, the paper would advance humanoid control beyond pre-scripted choreography by making skill selection responsive to live audio, with a modular pipeline that transfers from simulation to a physical robot. The paper is transparent about the transition bottleneck in the 20 s mashup and provides supplementary materials, which is commendable. However, the central evidence for 'semantic audio-driven understanding' rests on retrieval consistency with the authors' own timed skill rules on the same four reference tracks; the semantic-generalization component (CLAP fallback, novel audio) is not evaluated. The result is therefore currently an engineering demonstration of lookup-based orchestration rather than a validated semantic mapping.

major comments (4)
  1. [Section 4 / Section 3 (timed skill rules)] The ground truth for the 84.8% accuracy is the authors' hand-authored timed skill rules r=(s*, [t_start,t_end), π) defined in Section 3. A chunk is correct only if the retrieved policy matches the policy the authors assigned to that fragment of the same four songs. This protocol measures consistency with a hand-coded lookup table, not semantic appropriateness or generalization. The central claim of 'semantic' understanding is therefore untested by this metric. Please add evaluation on out-of-library tracks, independent human annotations of appropriate policies, or ablations showing that the rules are not arbitrary.
  2. [Section 3 (music retrieval) / Section 4] The CLAP embedding fallback is the only component capable of supporting the claimed generalization to novel audio inputs (Section 5), but it is never evaluated separately. All 574 test chunks are generated from the four reference tracks, so the Wang fingerprint path alone—which is an exact-match look-up—can solve the benchmark. Please report the fraction of chunks resolved by the fallback, measure its retrieval accuracy on tracks absent from the index, and include at least one out-of-library audio condition in the quantitative evaluation.
  3. [Section 4 (experimental protocol)] The single aggregate 84.8% figure is not statistically interpretable as reported. No confidence intervals, per-song or per-offset breakdown, or baseline comparisons (e.g., majority class, random, or a simple classifier) are given. Additionally, the hand-tuned routing thresholds c_min, v_min, v_frac_min, epsilon_sp, and the cooldown window are not assigned numerical values in Section 3, which makes the experimental pipeline non-reproducible. Please provide these values and a breakdown of accuracy across conditions.
  4. [Section 4.1 (M20, Figure 4)] The M20 Gantt chart shows that the controller frequently fails to complete transitions within 20-second sections and falls back to locomotion due to stability triggers. The paper candidly attributes this to transition latency of the stand-priming phase, but this limitation directly tempers the title-level claim of 'dynamic' whole-body control. The operating-point conclusion (30 s is suitable) should be accompanied by a quantitative measure of transition latency or an explicit scope statement restricting the claimed autonomy to the evaluated granularity.
minor comments (3)
  1. [Throughout] Typos and grammar: 'The remainder is organized as follow' should be 'as follows'; 'one of the example' should be 'one example'; 'represent a problem statement' should be 'is a problem'. Please proofread.
  2. [References [33]] Reference [33] for CLAP points to a multi-modal sentiment-analysis paper, not the commonly used audio-text CLAP model (LAION CLAP). Please cite the correct model or clarify the exact checkpoint used. Similarly, the proprietary 'gpt-4o-mini-transcribe' endpoint needs a version/date for reproducibility.
  3. [Section 4] The shift protocol is described only as an 'initial temporal offset of 0.5, 1.0, 1.5, and 2.0 seconds'. It is unclear whether each shifted chunk is a single 5 s window starting at that offset, how many chunks per song and per offset, and how the 574 total is composed. Please provide a per-condition breakdown.

Circularity Check

0 steps flagged

No circular derivation: retrieval accuracy is an empirical measurement against hand-authored skill rules, with no fitted parameter or self-citation load-bearing.

full rationale

The derivation chain is self-contained with respect to circularity. The reported 84.8% chunk-level accuracy measures the audio retrieval chain (Wang fingerprinting plus CLAP fallback, offset estimation, and timed-rule resolution) against ground-truth policies that are generated from the same timed skill rules R = (s*, [t_start, t_end), pi) described in Section 3. This shared definition means the benchmark validates consistency of track/offset lookup and policy resolution, not the external semantic validity of the hand-authored choreography; that is a validity limitation, but it is not a by-construction equivalence between a predicted quantity and a fitted input. No free parameter is fitted to force the reported accuracy, no result is claimed solely on a self-citation (the infrastructure citations RoboJuDo, BeyondMimic, CLAP, AST, etc. are prior external works), and no uniqueness theorem or ansatz is imported from the authors' prior work. The speech branch is presented only qualitatively and the CLAP-only/novel-audio generalization is untested, which limits the strength of the 'semantic understanding' claim and is acknowledged as needing further validation in the conclusion, but these are evaluation gaps rather than circular reasoning.

Axiom & Free-Parameter Ledger

7 free parameters · 5 axioms · 0 invented entities

The system is built entirely from existing pre-trained models, RL frameworks, and hand-authored rules. No new physical or conceptual entities are introduced. The main hidden load is a set of unreported thresholds and the hand-coded segment-to-skill mapping that defines ground truth.

free parameters (7)
  • chunk duration T
    Fixed at 5 s for all tests; affects temporal alignment and retrieval statistics; no sensitivity analysis is reported.
  • fingerprint acceptance confidence threshold c_min
    Minimum confidence score for accepting a fingerprint match; numerical value not reported.
  • aligned landmark votes threshold v_min
    Minimum number of aligned landmark votes for accepting a match; numerical value not reported.
  • speech VAD fraction threshold v_frac_min
    A chunk is routed to speech when the voiced frame fraction exceeds this threshold; value not reported.
  • speech-music margin epsilon_sp
    Speech branch requires p_speech - p_music >= epsilon_sp; value not reported.
  • cooldown window duration
    Suppresses repeated emissions of the same skill identifier; duration not reported.
  • timed skill rules R
    Hand-authored mapping (s*, [t_start,t_end), pi) from track segments to skills; defines the ground truth used in evaluation and is not independently validated.
axioms (5)
  • domain assumption Pre-trained AST on AudioSet classifies music vs speech sufficiently for routing
    The router sums posteriors over 527 AudioSet classes to get p_music and p_speech; no fine-tuning or domain adaptation is reported.
  • domain assumption CLAP embedding space aligns audio and text semantics well enough for top-1 skill retrieval
    Fallback retrieval uses cosine similarity in CLAP space; the paper only tests this implicitly through the four-track benchmark.
  • domain assumption MuJoCo simulation of the Unitree G1 is a faithful proxy for the physical robot
    Real-world deployment substitutes the simulated environment with hardware while keeping the control logic unchanged; sim-to-real fidelity is asserted, not measured.
  • domain assumption The hand-authored timed skill rules are semantically correct ground truth
    Evaluation counts a chunk correct if the retrieved policy matches the policy in the authors' rules; the semantic validity of those rules is not independently established.
  • domain assumption External speech-to-text and LLM APIs are reliable and low-latency in live operation
    The speech branch depends on OpenAI gpt-4o-mini-transcribe and gpt-4o-mini; latency and reliability are acknowledged as issues but not benchmarked.

pith-pipeline@v1.3.0-alltime-deepseek · 7763 in / 10939 out tokens · 101125 ms · 2026-08-02T03:25:59.364280+00:00 · methodology

0 comments
read the original abstract

Recent advances in humanoid robotics and reinforcement learning have enabled the acquisition of highly expressive whole-body motion policies. However, most robotic performances remain based on pre-scripted sequences or externally triggered behaviors, limiting autonomy and responsiveness to dynamic environments. In this work, we introduce a novel multi-modal orchestration framework for semantic audio-driven humanoid control, enabling robots to autonomously select and execute appropriate motion skills in real time. The system processes continuous audio streams and routes them into music or speech branches. Music input is handled via audio fingerprinting and semantic embeddings to retrieve track identity and temporal alignment, allowing dynamic mapping between musical segments and motion policies. Speech input is grounded into a discrete library of imitation-learned skills, enabling direct human-robot interaction. Both modalities share a unified interface that schedules skill execution over a reinforcement learning control pipeline. We validate the approach in simulation and on a Unitree G1 humanoid, showing robust sim-to-real transfer and consistent audio-conditioned policy selection. Supplementary materials are available at the following site: https://lab-rococo-sapienza.github.io/semantic-WBC/

Figures

Figures reproduced from arXiv: 2607.14182 by D. D. Bloisi, D. Nardi, E. Bugli, J. M. A. Marcelo, L. Comito, M. Brienza, V. Suriani.

Figure 1
Figure 1. Figure 1: Example of human motion retargeting in simulation and real-world de [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Proposed architecture. Continuous microphone audio is segmented into fixed-length chunks, each chunk is assigned a skill identifier by a routing and retrieval chain, and the chosen skill is executed on the humanoid through a switching layer that mediates transitions. 3 Methodology The proposed method shown in [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Sim-to-sim execution in the RoboJuDo [34] environment. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Gantt chart for the 20-second mashup (M20). The controller struggles to [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Gantt chart for the 30-second mashup (M30). Policy transitions remain [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Real-world execution on the Unitree G1, where music recognition drives [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

34 extracted references · 5 linked inside Pith

  1. [1]

    arXiv preprint arXiv:2510.02252 (2025) Semantic Audio for Dynamic Humanoid Control 11

    Araujo,J.P.,Ze,Y.,Xu,P.,Wu,J.,Liu,C.K.:Retargetingmatters:Generalmotion retargeting for humanoid motion tracking. arXiv preprint arXiv:2510.02252 (2025) Semantic Audio for Dynamic Humanoid Control 11

    arXiv 2025

  2. [2]

    arXiv preprint arXiv:2402.16796 (2024)

    Cheng, X., Ji, Y., Chen, J., Yang, R., Yang, G., Wang, X.: Expressive whole-body control for humanoid robots. arXiv preprint arXiv:2402.16796 (2024)

    Pith/arXiv arXiv 2024

  3. [3]

    In: Robotics: Science and Systems (RSS) (2024)

    Cheng, X., Ji, Y., Chen, J., Yang, R., Yang, G., Wang, X.: Expressive whole-body control for humanoid robots. In: Robotics: Science and Systems (RSS) (2024)

    2024

  4. [4]

    In: Conference on Robot Learning (CoRL) (2024)

    Fu, Z., Zhao, Q., Wu, Q., Wetzstein, G., Finn, C.: Humanplus: Humanoid shadow- ing and imitation from humans. In: Conference on Robot Learning (CoRL) (2024)

    2024

  5. [5]

    In: 2024 IEEE International Conference on Robotics and Automation (ICRA)

    Gao, J., Sarkar, B., Xia, F., Xiao, T., Wu, J., Ichter, B., Majumdar, A., Sadigh, D.: Physically grounded vision-language models for robotic manipulation. In: 2024 IEEE International Conference on Robotics and Automation (ICRA). pp. 12462– 12469. IEEE (2024)

    2024

  6. [6]

    2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) pp

    Gemmeke, J.F., Ellis, D.P.W., Freedman, D., Jansen, A., Lawrence, W., Moore, R.C., Plakal, M., Ritter, M.: Audio set: An ontology and human-labeled dataset for audio events. 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) pp. 776–780 (2017)

    2017

  7. [7]

    arXiv preprint arXiv:2104.01778 (2021)

    Gong, Y., Chung, Y.A., Glass, J.: Ast: Audio spectrogram transformer. arXiv preprint arXiv:2104.01778 (2021)

    Pith/arXiv arXiv 2021

  8. [8]

    In: 2024 IEEE 36th International Conference on Tools with Artificial Intelligence (ICTAI)

    Guan, H., Wei, X., Long, W., Yang, D., Zhai, P., Zhang, L.: Mdrpc: Music-driven robot primitives choreography. In: 2024 IEEE 36th International Conference on Tools with Artificial Intelligence (ICTAI). pp. 748–755. IEEE (2024)

    2024

  9. [9]

    In: Conference on Robot Learning (CoRL) (2024)

    He,T.,Luo,Z.,He,X.,Xiao,W.,Zhang,C.,Zhang,W.,Kitani,K.,Liu,C.,Shi,G.: Omnih2o: Universal and dexterous human-to-humanoid whole-body teleoperation and learning. In: Conference on Robot Learning (CoRL) (2024)

    2024

  10. [10]

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

    He, T., Luo, Z., Xiao, W., Zhang, C., Kitani, K., Liu, C., Shi, G.: Learning human- to-humanoid real-time whole-body teleoperation. In: 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 8944–8951. IEEE (2024)

    2024

  11. [11]

    Science robotics 4(26), eaau5872 (2019)

    Hwangbo,J.,Lee,J.,Dosovitskiy,A.,Bellicoso,D.,Tsounis,V.,Koltun,V.,Hutter, M.: Learning agile and dynamic motor skills for legged robots. Science robotics 4(26), eaau5872 (2019)

    2019

  12. [12]

    In: Proceedings of the Seventh Inter- national Conference on Computational Creativity

    Infantino, I., Augello, A., Manfré, A., Pilato, G., Vella, F.: Robodanza: Live per- formances of a creative dancing humanoid. In: Proceedings of the Seventh Inter- national Conference on Computational Creativity. pp. 388–395 (2016)

    2016

  13. [13]

    In: 2024 33rd IEEE International Conference on Robot and Human Interactive Communication (ROMAN)

    Ingelhag,N.,Munkeby,J.,vanHaastregt,J.,Varava,A.,Welle,M.C.,Kragic,D.:A robotic skill learning system built upon diffusion policies and foundation models. In: 2024 33rd IEEE International Conference on Robot and Human Interactive Communication (ROMAN). pp. 748–754. IEEE (2024)

    2024

  14. [14]

    arXiv preprint arXiv:2601.12799 (2026)

    Li, P., Zhuang, Z., Gao, Y., Dong, Y., Li, S., Jiang, C., Dou, S., Xi, Z., Zhou, E., Huang, J., et al.: From-w1: Towards general humanoid whole-body control with language instructions. arXiv preprint arXiv:2601.12799 (2026)

    arXiv 2026

  15. [15]

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

    Li, R., Yang, S., Ross, D.A., Kanazawa, A.: Ai choreographer: Music conditioned 3d dance generation with aist++. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 13401–13412 (2021)

    2021

  16. [16]

    arXiv preprint arXiv:2512.23650 (2026)

    Li, Z., Chi, C., Wei, Y., Zhu, B., et al.: Do you have freestyle? expressive humanoid locomotion via audio control. arXiv preprint arXiv:2512.23650 (2026)

    arXiv 2026

  17. [17]

    arXiv preprint arXiv:2508.08241 (2025)

    Liao, Q., Truong, T.E., Huang, X., Gao, Y., Tevet, G., Sreenath, K., Liu, C.K.: Beyondmimic: From motion tracking to versatile humanoid control via guided dif- fusion. arXiv preprint arXiv:2508.08241 (2025)

    Pith/arXiv arXiv 2025

  18. [18]

    McCarthy, S., He, F.: China’s biggest TV event had a clear star: the robot (2 2026), https://edition.cnn.com/2026/02/18/china/china-humanoid-robots-new-y ear-gala-intl-hnk 12 Marcelo et al

    2026

  19. [19]

    In: 2022 International Conference on Robotics and Automation (ICRA)

    Nie, B., Gao, Y.: Dancehat: generate stable dances for humanoid robots with ad- versarial training. In: 2022 International Conference on Robotics and Automation (ICRA). pp. 8511–8517. IEEE (2022)

    2022

  20. [20]

    ACM Transactions On Graphics (TOG)37(4), 1–14 (2018)

    Peng, X.B., Abbeel, P., Levine, S., Van de Panne, M.: Deepmimic: Example-guided deep reinforcement learning of physics-based character skills. ACM Transactions On Graphics (TOG)37(4), 1–14 (2018)

    2018

  21. [21]

    ACM Transactions on Graphics (ToG)40(4), 1–20 (2021)

    Peng, X.B., Ma, Z., Abbeel, P., Levine, S., Kanazawa, A.: Amp: Adversarial motion priors for stylized physics-based character control. ACM Transactions on Graphics (ToG)40(4), 1–20 (2021)

    2021

  22. [22]

    arXiv preprint arXiv:2410.03654 (2024)

    Radosavovic, I., Kamat, S., Darrell, T., Malik, J.: Learning humanoid locomotion over challenging terrain. arXiv preprint arXiv:2410.03654 (2024)

    Pith/arXiv arXiv 2024

  23. [23]

    Science Robotics9(89), eadi9579 (2024)

    Radosavovic, I., Xiao, T., Zhang, B., Darrell, T., Malik, J., Sreenath, K.: Real- world humanoid locomotion with reinforcement learning. Science Robotics9(89), eadi9579 (2024)

    2024

  24. [24]

    In: SIGGRAPH Asia 2024 Conference Papers

    Shen,Z.,Pi,H.,Xia,Y.,Cen,Z.,Peng,S.,Hu,Z.,Bao,H.,Hu,R.,Zhou,X.:World- grounded human motion recovery via gravity-view coordinates. In: SIGGRAPH Asia 2024 Conference Papers. pp. 1–11 (2024)

    2024

  25. [25]

    In: Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Siyao, L., Yu, W., Gu, T., Lin, C., Wang, Q., Qian, C., Loy, C.C., Liu, Z.: Bailando: 3d dance generation by actor-critic gpt with choreographic memory. In: Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11050–11059 (2022)

    2022

  26. [26]

    arXiv preprint arXiv:2509.20717 (2025)

    Sun, Z., Peng, Y., Meng, Y., Li, X., Huang, B.S., Bing, Z., Wang, X., Knoll, A.: Robotdancing: Residual-action reinforcement learning enables robust long-horizon humanoid motion tracking. arXiv preprint arXiv:2509.20717 (2025)

    arXiv 2025

  27. [27]

    Team, S.: Silero vad: pre-trained enterprise-grade voice activity detector (vad), number detector and language classifier.https://github.com/snakers4/silero -vad(2024), gitHub repository

    2024

  28. [28]

    In: 2012 IEEE/RSJ International Conference on Intelligent Robots and Systems

    Todorov, E., Erez, T., Tassa, Y.: Mujoco: A physics engine for model-based control. In: 2012 IEEE/RSJ International Conference on Intelligent Robots and Systems. pp. 5026–5033 (2012). https://doi.org/10.1109/IROS.2012.6386109

    arXiv 2012

  29. [29]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recog- nition

    Tseng, J., Castellon, R., Liu, K.: Edge: Editable dance generation from music. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recog- nition. pp. 448–458 (2023)

    2023

  30. [30]

    In: ISMIR 2003, 4th Symposium Conference on Music Information Retrieval

    Wang, A.L.: An industrial-strength audio search algorithm. In: ISMIR 2003, 4th Symposium Conference on Music Information Retrieval. pp. 7–13 (2003)

    2003

  31. [31]

    arXiv preprint arXiv:2603.15603 (2026)

    Yang, T., He, S., Jing, H., Yang, J., Liu, Z., Zou, C., Wang, Y.: Fast sam 3d body: Accelerating sam 3d body for real-time full-body human mesh recovery. arXiv preprint arXiv:2603.15603 (2026)

    arXiv 2026

  32. [32]

    arXiv preprint arXiv:2505.02833 (2025)

    Ze, Y., Chen, Z., Araújo, J.P., Cao, Z.a., Peng, X.B., Wu, J., Liu, C.K.: Twist: Tele- operated whole-body imitation system. arXiv preprint arXiv:2505.02833 (2025)

    Pith/arXiv arXiv 2025

  33. [33]

    In: Proceedings of the 2023 ACM International Conference on Multi- media Retrieval

    Zhao, T., Kong, M., Liang, T., Zhu, Q., Kuang, K., Wu, F.: Clap: Con- trastive language-audio pre-training model for multi-modal sentiment analy- sis. In: Proceedings of the 2023 ACM International Conference on Multi- media Retrieval. p. 622–626. Association for Computing Machinery (2023). https://doi.org/10.1145/3591106.3592296,https://doi.org/10.1145/35...

    arXiv 2023

  34. [34]

    just deploy, just do

    Zihan Zhuang, Yi Dong, P.L.: A plug-and-play deploy framework for robots. just deploy, just do. (2025),https://github.com/HansZ8/RoboJuDo

    2025