Pith. sign in

REVIEW 4 major objections 6 minor 25 references

Event cameras beat RGB for teleoperation under low light and backlit scenes, with half the latency and jitter.

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

2026-08-03 11:10 UTC pith:ET5ZSLSB

load-bearing objection Honest, well-scoped empirical paper on event-based teleoperation with a real baseline caveat that the authors already acknowledge. the 4 major comments →

arxiv 2607.29227 v1 pith:ET5ZSLSB submitted 2026-07-31 cs.RO

Event-Based Upper-Body Humanoid Teleoperation Under Challenging Illumination

Haoyu Fu , Zhou Ge , Chengze Li , Chenzhao Sun , Ze Cui , Wenjing Zhou , Xulei Qin This is my paper
classification cs.RO
keywords event camerahumanoid teleoperationhuman pose estimationchallenging illuminationhigh dynamic rangemotion retargetingreal-time controlembedded robotics
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.

This paper tries to establish that an event-camera pipeline can outperform standard RGB cameras for real-time upper-body humanoid teleoperation in challenging conditions—severe backlighting, low light below 5 lux, and fast arm motions—on the metrics that matter for control: latency, temporal jitter, lost-frame ratio, and commanded joint error. It constructs a complete embedded system that accumulates events into time-surfaces, estimates 3D pose with a lightweight network, fuses a gravity-aligned IMU, and retargets causally to an 18-DoF robot, achieving 23–34 ms end-to-end latency. In paired trials, the event pipeline shows roughly half the latency and jitter of a fixed-exposure RGB baseline and higher tracking success under stress, while RGB keeps a slight edge in frame-wise pose accuracy (MPJPE 31.7 vs 33.8 mm). A sympathetic reader should care because the paper frames the gain as closed-loop robustness rather than absolute pose accuracy, which is the property that actually determines teleoperation safety and operator immersion.

Core claim

The central claim is that for upper-body human-to-humanoid motion imitation, event-based sensing yields control-oriented advantages over frame-based RGB under adverse illumination and rapid motion: end-to-end photon-to-action latency of 27.4±3.6 ms vs 52.8±9.8 ms, temporal pose jitter of 10.8±3.1 mm vs 18.9±5.5 mm, robot joint RMSE of 4.9±1.3° vs 6.2±2.0°, and tracking success of 93.8% vs 38.6% under HDR backlight. These gains come at a small cost: RGB remains slightly better on per-frame 3D pose error (MPJPE 31.7 vs 33.8 mm). The paper's conclusion is a trade-off, not a universal win: events are preferable for fast or poorly lit teleoperation, while well-lit static scenes may favor RGB or a

What carries the argument

The argument is carried by three components working together: (1) a time-surface representation that accumulates asynchronous brightness-change events into a compact tensor over a 5 ms window with exponential decay, preserving microsecond timing while feeding a standard convolutional pose network; (2) a gravity-aligned inertial fusion plus a speed-based low-pass filter that suppresses temporal jitter; and (3) a causal kinematic retargeting optimizer that maps estimated human keypoints to robot joint commands under joint, velocity, and acceleration limits, down-weighting low-confidence joints and regularizing toward a nominal posture. The load-bearing mechanism is the coupling of high-tempora

Load-bearing premise

The comparison treats a fixed-exposure RGB camera with auto-gain disabled as the representative frame-based pipeline; if a well-tuned auto-exposure RGB system narrows or eliminates the event advantage, the central 'events beat RGB' claim is overstated.

What would settle it

Run the identical 12-subject teleoperation protocol under HDR backlight and low light with an RGB camera using per-frame auto-exposure and auto-gain, keeping the same downstream pose and retargeting stack. If the paired differences in photon-to-action latency (27.4 vs 52.8 ms), temporal jitter (10.8 vs 18.9 mm), and robot joint RMSE (4.9° vs 6.2°) shrink below significance, the paper's central claim would be falsified.

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

If this is right

  • Low-light or backlit teleoperation need not depend on RGB sensing; event cameras alone can carry the closed loop, with latency under 34 ms and joint error under 5°.
  • Evaluation of humanoid teleoperation should include control-oriented metrics—latency, jitter, lost frames, joint RMSE—because MPJPE alone misses exactly the failure modes that matter for safety.
  • Event+RGB hybrid systems are a natural next step: events handle fast motion and HDR, RGB provides absolute pose in static scenes where events go silent.
  • The lower perception power draw of event sensing (3.1 W vs 4.0-7.8 W for RGB) extends battery life for untethered robots.
  • Fast gestures up to 5 Hz can be tracked without frame-rate aliasing, enabling more natural high-speed interaction.

Where Pith is reading between the lines

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

  • The same control-vs-accuracy trade-off likely generalizes to other closed-loop tasks—drone teleoperation, surgical robotics, AR interaction—wherever latency and smoothness dominate over single-frame accuracy; the paper's evidence suggests such tasks might prefer event input even at a pose-accuracy cost.
  • The confidence-adaptive retargeting hints at a broader design principle: instead of perfecting perception, the control layer can explicitly soften uncertain joints, so robustness can be bought at the control level rather than in the vision network.
  • A direct testable extension: run the same protocol under well-tuned auto-exposure RGB; if the event advantage shrinks to insignificance, the practical claim reduces to 'events win against fixed-exposure sensors,' which is a narrower but still useful result. The paper itself acknowledges this possibility.
  • Another testable extension: measure operator task performance—completion time, error rate, subjective telepresence—under the same stress conditions; because human-in-the-loop feel is dominated by latency, the event advantage may be even larger than the kinematic metrics suggest.

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 / 6 minor

Summary. The paper presents an event-camera-based upper-body human-to-humanoid teleoperation system. Events are accumulated into a time-surface representation, fed into a compact 3D pose network, fused with IMU gravity alignment, filtered with One-Euro, and mapped to robot joint commands via a TWIST-style causal optimizer. The main experimental claim is that, under the authors' setup (12 subjects x 5 repetitions x 4 conditions), the event pipeline offers control-oriented advantages over a fixed-exposure RGB baseline: lower photon-to-action latency (27.4 vs 52.8 ms), lower temporal jitter, lower joint RMSE, and higher tracking success under HDR backlight, low light, and fast motion. The paper honestly reports that RGB retains an MPJPE advantage and explicitly frames the result as a trade-off rather than universal superiority. It also includes paired and mixed-effects statistical analyses, a latency breakdown, failure-recovery metrics, and a clearly written threats-to-validity section.

Significance. If the comparison holds under a better-matched baseline, this is a significant systems contribution. The paper's strengths include: paired experimental design with both trial-level and aggregated sensitivity analyses; use of control-relevant metrics (latency, jitter, joint RMSE, tracking success) rather than MPJPE alone; honest reporting of the RGB MPJPE advantage and the trade-off; explicit limitations in Section V-L; and a useful latency stage-by-stage breakdown. The paper does not oversell the event advantage as universal. However, the central empirical comparison rests on an RGB baseline with fixed exposure and disabled auto-gain, the event hyperparameters are fixed without sensitivity analysis, and no code/data/model weights are released; these issues limit the strength of the current claims.

major comments (4)
  1. [§V-B, Tables II–III, Abstract] The central claim of 'advantages over RGB baselines' is established only against an RGB baseline with fixed exposure and disabled auto-gain. This is a deliberate stress configuration, but it suppresses the standard adaptive behavior of commercial RGB cameras. The 93.8% vs 38.6% tracking-success gap in HDR backlight (Table III) could substantially reflect fixed foreground underexposure rather than an inherent sensor advantage. The paper's own caveat in §V-B that 'well-tuned auto-exposure could narrow some gaps' is load-bearing and should be tested or elevated. Either add an auto-exposure/tuned RGB baseline as a secondary condition, or consistently narrow all headline claims to 'fixed-exposure RGB comparison.' The latency advantage in Table IV is more robust, but the jitter/RMSE/tracking-success claims are not.
  2. [§V-D, Eq. (6), Table II] The mixed-effects model is incompletely specified. The text says 120 FPS is 'retained as a covariate,' and Eq. (6) includes a 1_{120FPS} term, but Table II reports only the event coefficient and p-value. It is not stated how many observations entered the LME: if the 120 FPS RGB trials are included, the analysis has 720 observations per metric; if not, the covariate is undefined. The phrase '240 paired condition-level observations per modality' is also confusing because 12x4x5 with repetitions is trial-level, not condition-level after aggregation. Please report the full model specification, sample sizes, random-effect variance estimates, and the coefficient for 120 FPS.
  3. [§III-C, §IV-C, §V-J] The event pipeline's hyperparameters—accumulation window Δt=5 ms, stride=1 ms, time-surface decay τ=5 ms, One-Euro β/f_min, and TWIST weights—are fixed across all experiments with no reported sensitivity analysis. Section V-J ablates the time-surface vs voxel representation, the One-Euro filter, IMU fusion, and confidence weighting, but not Δt/τ or the filter/optimizer parameters. Because the RGB baseline is deliberately left untuned, the comparison must not become a tuned-event vs untuned-RGB contrast. Please provide a sensitivity study for the most critical event parameters (at least Δt and τ) or justify their fixed values from a validation procedure.
  4. [§V-E, Table III] The robustness metrics that support the central claim—tracking success and lost-frame ratio—are reported only as means ± SD with no paired tests, confidence intervals, or effect sizes. The paper's statistical machinery is concentrated on Table II metrics. Since Table III contains the main evidence for 'advantages under stress,' these comparisons need inferential support or an explicit description as descriptive pilot results.
minor comments (6)
  1. [§V-D] The 95% bootstrap confidence intervals for MPJPE and joint RMSE are mentioned in the text but never reported in Table II or elsewhere. Please include them or remove the sentence.
  2. [§III-C, Eq. (1)] The time-surface equation does not define the polarity sign convention or whether the accumulated tensor is normalized before being fed to the network. Please specify the preprocessing exactly.
  3. [§V-C, Eq. (5)] Joint RMSE uses a 'reference value from synchronized demonstrations.' Please clarify how the demonstrations were synchronized across modalities and whether the same reference sequence was used for event and RGB trials.
  4. [§IV-A] The platform is named 'NVIDIA Booster T1.' I am not aware of an official NVIDIA product with that name; if it is a custom or renamed module, please provide the underlying SoC and power/thermal specifications.
  5. [§V-D] Multiple metrics are tested without multiple-comparison correction. This may be acceptable for an exploratory systems study, but it should be stated explicitly.
  6. [Throughout] No code, trained model weights, or dataset release is mentioned despite the 'Reproducibility Settings' section. For a systems paper of this type, releasing the perception model and evaluation scripts would materially improve reproducibility.

Circularity Check

0 steps flagged

No circularity: the event-vs-RGB comparison is an empirical measurement with explicit scoping and no fitted parameter renamed as a prediction.

full rationale

The paper's central claim—event-based sensing yields control-oriented advantages under challenging illumination—rests on direct experimental measurements (latency, jitter, joint RMSE, tracking success) rather than on a derivation whose output is equivalent to its input. The pose estimator is trained on external datasets (V2E-converted Human3.6M and DHP19) and evaluated on held-out event data; the reported MPJPE is actually slightly worse for events than RGB (33.8 vs 31.7 mm), which is inconsistent with any circular design intended to force event superiority. The RGB baseline uses fixed exposure and disabled auto-gain, and the paper explicitly discloses this choice in Section V-B ('fixed exposure/gain gives repeatable stress tests and avoids hidden adaptation in backlight, but is conservative in moderate lighting') and Section V-L ('well-tuned auto-exposure could narrow some gaps'). The conclusions are consistently scoped to 'under our experimental setup' and 'under our protocol,' and the paper does not claim universal superiority. The motion retargeting module (TWIST) and prior event-HPE methods are cited from external groups, with no load-bearing self-citation chain: none of the authors' own prior work is invoked to justify the central result. The 120 FPS RGB condition is not hidden; it is retained as a covariate in the mixed-effects model in Equation 6. The identified weaknesses—baseline tuning, pilot-scale statistics, and external validity—are validity threats, not circular reasoning.

Axiom & Free-Parameter Ledger

7 free parameters · 5 axioms · 0 invented entities

No new physical entities, forces, or mediators are introduced; all components are existing hardware/software. The report rests on hand-set perception and control parameters and on assumptions about generalization of the pose backbone and fairness of the RGB baseline. No formal derivation is involved; all numbers are empirical.

free parameters (7)
  • Event accumulation window Δt = 5 ms
    Time-surface window chosen by hand; affects pose estimation latency and accuracy (Eq. 1).
  • Time-surface decay τ = 5 ms
    Exponential decay constant in Eq. 1; hand-set and held fixed, no sensitivity analysis.
  • Time-surface stride = 1 ms
    Accumulation step; combined with Δt determines effective pose update rate.
  • One-Euro filter parameters β and f_min = β=0.4, f_min=1.5 Hz
    Smoothing parameters chosen by hand; ablation says removing them increases jitter by 45%, so they materially affect the reported jitter advantage.
  • TWIST weights w_t, w_s, w_l = 1.0, 0.1, 0.01
    Retargeting objective weights; hand-set, favor tracking over smoothness, and affect joint RMSE comparison.
  • Safety buffer = 10 ms
    Added to all latency totals; arbitrary safety margin, affects absolute delay claims.
  • RGB exposure/gain settings = unspecified
    Fixed exposure and disabled auto-gain for RGB baseline; exact values not stated, but baseline behavior drives the comparison.
axioms (5)
  • domain assumption A 5 ms exponential time-surface (Eq. 1) retains sufficient pose information from the event stream for 100 Hz control.
    The whole perception module rests on this representation; no formal or empirical comparison against other representations is provided except a voxel-grid latency ablation.
  • domain assumption The MediaPipe-like 3D pose backbone trained on V2E-converted Human3.6M and fine-tuned on DHP19 generalizes to live EVK4 events in the tested lighting conditions.
    No real-event pose accuracy numbers are reported for the final deployment conditions; only a held-out validation MPJPE of 31.8 mm is stated.
  • domain assumption TWIST retargeting (Eq. 2) with weights w_t=1.0, w_s=0.1, w_l=0.01 maps human poses to safe robot commands at 100 Hz.
    The objective is taken from prior work [25] and its weights are hand-set; no sensitivity analysis is provided for these weights.
  • domain assumption The RGB baseline with fixed exposure and disabled auto-gain represents a fair comparison for standard frame-based pipelines.
    The authors acknowledge auto-exposure could narrow the gap; if the baseline is not representative of deployed RGB systems, the relative advantage is overstated.
  • domain assumption The 18-DoF humanoid platform and the Booster T1 are representative enough for upper-body teleoperation metrics.
    Experiments use one robot platform, upper-body only; external validity is limited as acknowledged in Section V-L.

pith-pipeline@v1.3.0-daily-deepseek · 9935 in / 12211 out tokens · 109443 ms · 2026-08-03T11:10:59.200225+00:00 · methodology

0 comments
read the original abstract

We present a real-time upper-body human-to-humanoid motion imitation framework driven by neuromorphic event-based vision. This work addresses practical perceptual bottlenecks of conventional frame-based RGB sensors, specifically their difficulty in high dynamic range (HDR) scenes and rapid motions due to fixed integration times. By leveraging the Prophesee EVK4 event camera, which operates asynchronously with high temporal resolution and a dynamic range exceeding 120 dB, our system supports stable tracking in conditions where standard vision pipelines degrade, such as severe backlighting and very low light environments below 5 lux. The architecture integrates a low-latency Perception Module, utilizing optimized event accumulation and gravity-aligned inertial fusion, with a causal Motion Module (TWIST) that performs online kinematic retargeting. We validate the system on an embedded NVIDIA Booster T1 platform and an 18-DoF humanoid upper-body setup, demonstrating an end-to-end photon-to-action latency of 23-34 ms and advantages over RGB baselines under our experimental setup. The results indicate a practical trade-off: events can be preferable for fast or poorly lit upper-body teleoperation, whereas well-lit static scenes may favor RGB or hybrid sensing.

Figures

Figures reproduced from arXiv: 2607.29227 by Chengze Li, Chenzhao Sun, Haoyu Fu, Wenjing Zhou, Xulei Qin, Ze Cui, Zhou Ge.

Figure 1
Figure 1. Figure 1: Event-based human-to-robot imitation system overview. Events are [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Experimental setup for the qualitative robot-imitation results in [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparison under fast motions with labeled subfigures [PITH_FULL_IMAGE:figures/full_fig_p006_3.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

25 extracted references · 2 linked inside Pith

  1. [1]

    End-to-end recovery of human shape and pose,

    A. Kanazawa, M. J. Black, D. W. Jacobs, and J. Malik, “End-to-end recovery of human shape and pose,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 7122–7131

    2018

  2. [2]

    Synthesis of whole-body behaviors through hierarchical control of behavioral primitives,

    L. Sentis and O. Khatib, “Synthesis of whole-body behaviors through hierarchical control of behavioral primitives,”International Journal of Humanoid Robotics, vol. 4, no. 04, pp. 741–771, 2007

    2007

  3. [3]

    Event-based vision: A survey,

    G. Gallego, T. Delbr ¨uck, G. Orchard, C. Bartolozzi, B. Taba, A. Censi, S. Leutenegger, A. J. Davison, J. Conradt, K. Daniilidiset al., “Event-based vision: A survey,”IEEE transactions on pattern analysis and machine intelli- gence, vol. 44, no. 1, pp. 154–180, 2020

    2020

  4. [4]

    A 128×128 120 db 15µs latency asynchronous temporal contrast vision sensor,

    P. Lichtsteiner, C. Posch, and T. Delbruck, “A 128×128 120 db 15µs latency asynchronous temporal contrast vision sensor,”IEEE journal of solid- state circuits, vol. 43, no. 2, pp. 566–576, 2008

    2008

  5. [5]

    A 240×180 130 db 3µs latency global shutter spatiotemporal vision sensor,

    C. Brandli, R. Berner, M. Yang, S.-C. Liu, and T. Delbruck, “A 240×180 130 db 3µs latency global shutter spatiotemporal vision sensor,”IEEE Journal of Solid-State Circuits, vol. 49, no. 10, pp. 2333–2341, 2014

    2014

  6. [6]

    Ultimate SLAM? combining events, images, and IMU for robust visual SLAM in HDR and high- speed scenarios,

    A. R. Vidal, H. Rebecq, T. Horstschaefer, and D. Scaramuzza, “Ultimate SLAM? combining events, images, and IMU for robust visual SLAM in HDR and high- speed scenarios,”IEEE Robotics and Automation Letters, vol. 3, no. 2, pp. 994– 1001, 2018

    2018

  7. [7]

    EventCap: Monocular 3d capture of high-speed human motions using an event camera,

    L. Xu, W. Xu, V . Golyanik, M. Habermann, L. Fang, and C. Theobalt, “EventCap: Monocular 3d capture of high-speed human motions using an event camera,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020, pp. 4968–4978

    2020

  8. [8]

    SMPL: A skinned multi-person linear model,

    M. Loper, N. Mahmood, J. Romero, G. Pons-Moll, and M. J. Black, “SMPL: A skinned multi-person linear model,”ACM transactions on graphics (TOG), vol. 34, no. 6, pp. 1–16, 2015

    2015

  9. [9]

    Deepmimic: Example- guided deep reinforcement learning of physics-based character skills,

    X. B. Peng, P. Abbeel, S. Levine, and M. Van de Panne, “Deepmimic: Example- guided deep reinforcement learning of physics-based character skills,”ACM Transactions On Graphics (TOG), vol. 37, no. 4, pp. 1–14, 2018

    2018

  10. [10]

    Retargeting matters: General mo- tion retargeting for humanoid motion tracking,

    J. P. Araujo, Y . Ze, P. Xu, J. Wu, and C. K. Liu, “Retargeting matters: General mo- tion retargeting for humanoid motion tracking,”arXiv preprint arXiv:2510.02252, 2025

    arXiv 2025

  11. [11]

    Retargeting motion to new characters,

    M. Gleicher, “Retargeting motion to new characters,” inProceedings of the 25th annual conference on Computer graphics and interactive techniques, 1998, pp. 33–42

    1998

  12. [12]

    Motion retargeting for humanoid robots based on identification to preserve and reproduce human motion features,

    K. Ayusawa and E. Yoshida, “Motion retargeting for humanoid robots based on identification to preserve and reproduce human motion features,” in2014 IEEE/RSJ International Conference on Intelligent Robots and Systems. IEEE, 2014, pp. 4464–4469

    2014

  13. [13]

    Robust real-time whole-body motion retargeting from human to humanoid,

    L. Penco, B. Cl ´ement, V . Modugno, E. M. Hoffman, G. Nava, D. Pucci, N. G. Tsagarakis, J.-B. Mouret, and S. Ivaldi, “Robust real-time whole-body motion retargeting from human to humanoid,” in2018 IEEE-RAS 18th International Conference on Humanoid Robots (Humanoids). IEEE, 2018, pp. 425–432

    2018

  14. [14]

    ESIM: an open event camera simulator,

    H. Rebecq, D. Gehrig, and D. Scaramuzza, “ESIM: an open event camera simulator,” inProceedings of the 2nd Conference on Robot Learning (CoRL). PMLR, 2018, pp. 969–982

    2018

  15. [15]

    Dynamic obstacle avoidance for quadrotors with event cameras,

    D. Falanga, K. Kleber, and D. Scaramuzza, “Dynamic obstacle avoidance for quadrotors with event cameras,”Science Robotics, vol. 5, no. 40, p. eaaz9712, 2020

    2020

  16. [16]

    Eventhpe: Event-based 3D human pose and shape estimation,

    S. Zou, C. Guo, X. Zuo, S. Wang, P. Wang, X. Hu, S. Wu, and L. Cheng, “Eventhpe: Event-based 3D human pose and shape estimation,” inProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021, pp. 11 613–11 622

    2021

  17. [17]

    End-to-end learning of representations for asynchronous event-based data,

    D. Gehrig, A. Loquercio, K. G. Derpanis, and D. Scaramuzza, “End-to-end learning of representations for asynchronous event-based data,” inProceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 5633– 5643

    2019

  18. [18]

    Lifting monocular events to 3D human poses,

    G. Scarpellini, P. Morerio, and A. Del Bue, “Lifting monocular events to 3D human poses,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2021, pp. 1439–1449

    2021

  19. [19]

    EventEgo3D++: 3D human motion capture from a head-mounted event camera,

    C. Millerdurai, H. Akada, J. Wang, D. Luvizon, C. Theobalt, and V . Golyanik, “EventEgo3D++: 3D human motion capture from a head-mounted event camera,” International Journal of Computer Vision, 2025

    2025

  20. [20]

    DHP19: Dynamic vision sensor 3D human pose dataset,

    E. Calabrese, G. Taverni, C. E. Awai, D. Scaramuzzaet al., “DHP19: Dynamic vision sensor 3D human pose dataset,” inProceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition Workshops (CVPRW), 2019

    2019

  21. [21]

    HOTS: A hierarchy of event-based time-surfaces for pattern recognition,

    X. Lagorce, G. Orchard, F. Galluppi, B. E. Shi, and R. B. Benosman, “HOTS: A hierarchy of event-based time-surfaces for pattern recognition,”IEEE Transac- tions on Pattern Analysis and Machine Intelligence, vol. 39, no. 7, pp. 1346–1359, 2016

    2016

  22. [22]

    Mediapipe: A framework for building perception pipelines,

    C. Lugaresi, J. Tang, H. Nash, C. McClanahan, E. Uboweja, M. Hays, F. Zhang, C.-L. Chang, M. G. Yong, J. Leeet al., “Mediapipe: A framework for building perception pipelines,”arXiv preprint arXiv:1906.08172, 2019

    Pith/arXiv arXiv 1906

  23. [23]

    v2e: From video frames to realistic DVS events,

    Y . Hu, S.-C. Liu, and T. Delbruck, “v2e: From video frames to realistic DVS events,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 1312–1321

    2021

  24. [24]

    1 C filter: a simple speed-based low- pass filter for noisy input in interactive systems,

    G. Casiez, N. Roussel, and D. V ogel, “1 C filter: a simple speed-based low- pass filter for noisy input in interactive systems,” inProceedings of the SIGCHI Conference on human factors in computing systems, 2012, pp. 2527–2530

    2012

  25. [25]

    Twist: Teleoperated whole-body imitation system,

    Y . Ze, Z. Chen, J. P. Ara ´ujo, Z.-a. Cao, X. B. Peng, J. Wu, and C. K. Liu, “Twist: Teleoperated whole-body imitation system,”arXiv preprint arXiv:2505.02833, 2025

    Pith/arXiv arXiv 2025