Pith. sign in

REVIEW 4 major objections 5 minor 38 references

TOP: Time Optimization Policy for Stable and Accurate Standing Manipulation with Humanoid Robots

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper argues that adaptively slowing upper-body motions—not just strengthening lower-body stability—is what lets a standing humanoid stay both balanced and precise during manipulation tasks, and it reports 95.30% task success with…

desk verdict A genuine, well-tested idea — learned time-scaling of upper-body motions — but the feasibility guarantee is asserted, not proven, and the headline metrics exclude failure trials. read the letter →

arxiv 2508.00355 v1 pith:YCEHSWD4 submitted 2025-08-01 cs.RO

Zhenghan Chen , Haocheng Xu , Haodong Zhang , Liang Zhang , He Li , Dongqi Wang , Jiyu Yu , Yifei Yang
show 2 more authors
Zhongxiang Zhou Rong Xiong
This is my paper
classification cs.RO
keywords humanoidstandingmanipulationtimeoptimizationpolicywhole-bodycontrolupper-bodymotionpriorvariationalautoencoderreinforcementlearningbalanceretargeting
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

This paper argues that the right time to act—not just the right action—is a missing control variable in standing humanoid manipulation. It proposes TOP, a supervised reinforcement-learning policy that decides, at every step, how long to wait between successive upper-body motion clips, slowing the arms precisely when their momentum would disturb the lower-body balance policy. Combined with a variational-autoencoder motion prior and a decoupled controller (PD for the arms, reinforcement learning for the legs), TOP raises success to 95.30% in simulation tasks and reduces endpoint tracking error relative to whole-body RL baselines. If the claim holds, humanoid controllers can keep precision and stability without hand-tuning motion speed.

What carries the argument

The load-bearing mechanism is the Time Optimization Policy itself: a supervised RL actor-critic that maps the current motion clip $m_t$, its latent code $z_t$, the robot state $s_t$, and history $h_t$ to a chunk of future timestamps $\Delta t^{\mathrm{seq}}_t$. Output timestamps are applied by replacing the original motion with a linearly interpolated clip $m'_t = \mathrm{linear\,interpolate}(m_{t-1}, m_{t+\Delta t_t})$, and the policy is trained with a weighted combination of a supervised loss that anchors timestamps near the dataset's original spacing and a PPO loss that rewards balance, accuracy, and small smooth time steps. Action chunking with exponential weights smooths the sequence and prevents jerky switching. The VAE motion prior and the decoupled PD-plus-RL controller support TOP by giving the balance policy advance knowledge of upper-body motion.

What would settle it

Run a set of interpolated motion pairs through a feasibility audit: compute joint torque commands, velocity limits, and the zero-moment-line distance for every interpolated frame, and compare success rates with and without a feasibility filter. If many interpolated trajectories violate constraints or if filtering them removes most of TOP's advantage, the interpolation assumption is the weak link.

Watch

Extended reading notes

Core claim

The central claim is that adaptive time scaling of reference motions resolves the standing-manipulation trade-off between precision, stability, and time efficiency. In the paper's architecture, upper-body motion clips are represented as latent codes by a VAE, executed accurately by a PD controller, and stabilized by a lower-body RL policy; TOP then learns to output a sequence of timestamps $\Delta t_t,\ldots,\Delta t_{t+N}$ that slow down the current clip when fast arm motion would destabilize the robot, using linear interpolation between the previous and next clip. With this module, the robot reaches 95.30% success and end-effector position error 0.0270, compared with 92.46% and 0.0741 for Exbody and 94.08% and 0.0506 for OmniH2O baselines. The paper also reports that the same controller keeps the zero-moment line inside the support polygon and maintains comparable accuracy on a physical robot.

Load-bearing premise

The argument depends on linear interpolation between dataset motion clips always producing a new motion that still obeys the robot's kinematic and dynamic constraints, so the upper-body PD controller is never asked to track an infeasible reference.

Editorial extensions

If this is right

  • Reference motions from teleoperation, planners, or vision-language models can be executed without manually setting speed; TOP supplies timing online.
  • Hard motion segments are slowed only when needed, preserving time efficiency on easy segments and improving stability and precision on hard ones.
  • Because the lower-body controller and TOP are trained separately, swapping in a new lower-body policy requires retraining TOP rather than the whole stack.
  • The decoupled architecture with TOP can be extended from standing to loco-manipulation, as the paper's appendix demonstrates with walking manipulation tasks.
  • Unseen, faster motions are handled with over 80% success, indicating that the time-optimization strategy generalizes beyond the training motion set.

Reading between the lines

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

  • A direct test of the interpolation assumption would be to check interpolated trajectories against joint torque and velocity limits and monitor the zero-moment point; if violations appear, a safety filter could replace the current unverified assumption.
  • The same timestamp-optimization idea might apply to other whole-body tasks where momentum from one body segment disturbs another, for example coordinating arm swing with step phase in locomotion or with changing payloads.
  • The Pareto analysis in the appendix suggests TOP learns a point on a precision–stability–time frontier; making that frontier explicit would let users choose a time budget instead of accepting the learned default.
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 proposes TOP, a supervised reinforcement-learning module that adjusts the timestamps between upper-body motion clips for a humanoid standing-manipulation controller. The framework has three components: a VAE that encodes upper-body motion windows into a latent motion prior; a decoupled controller in which a lower-body RL policy maintains balance while an upper-body PD controller tracks the reference; and TOP, which outputs a horizon of delta-t adjustments that are applied by linearly interpolating between consecutive motion states. The authors evaluate the method in IsaacGym simulation and on a real 41-DoF humanoid over more than 10,000 motion clips, reporting a 95.30% success rate and lower upper-body joint and end-effector errors than whole-body RL and fixed-timestep decoupled baselines, at the cost of a larger time cost (40.5 s vs. 15.0 s nominal). Appendices add domain randomization, payload-adaptability tests, ZML stability analysis, momentum analysis, sensitivity studies, and a regression/Pareto analysis of the time-stability-precision trade-off.

Significance. If the results hold, the paper offers a practical and modular mechanism for time-scaling reference motions so that a humanoid can trade off stability, precision, and execution speed: the robot slows down only when fast upper-body motion would threaten balance. The strengths are the large-scale simulation test set, real-world validation, systematic ablations, domain randomization, and the unusually detailed appendix material on sensitivity, payload, and stability metrics. The main limitations are that several load-bearing claims rest on an unsupported feasibility assertion and on metrics computed over a favorable trial subset, so the empirical evidence needs to be strengthened before the paper's conclusions can be taken at face value.

major comments (4)
  1. [III-D] The assertion that linear interpolation preserves dynamic feasibility is unsupported and load-bearing. The text states: 'Because the motions of the dataset M satisfy the kinematic and dynamic constraints of the robot, after linear interpolation, m′_t will not violate the constraints.' This does not follow: while joint-limit constraints are convex, velocity and acceleration bounds, torque limits, and ZMP conditions are not preserved by linearly interpolating between states separated by up to 0.1 s (five 50 Hz frames). No proof, interpolation error bound, or constraint checker is provided. Appendix C explicitly concedes that the VAE is purely kinematic and that the downstream RL policy compensates for dynamic variations, which directly undercuts the blanket feasibility claim in III-D. Because the upper-body PD controller tracks m′_t, an infeasible reference would invalidate both the precision and stability benefits attributed to TOP; the authors should either prove a bound, add a feasibility filter, or empirically verify that interpolated reference states satisfy torque and ZMP limits.
  2. [IV, Table III] The mean performance metrics in Table III are computed only on trials where the robot remains standing in place, with steps and falls excluded by the table caption. The success rate, by contrast, is computed over all trials. This means the reported precision and stability values (e.g., E_eepe = 0.0270) average over a favorable subset, and the subset size differs across methods because TOP's mechanism is precisely to reduce falls and steps. Comparing conditional means across methods conflates the ability to avoid failure with the ability to control during successful trials. The paper should report metrics on all trials with an explicit convention (for example, a failure penalty) or at least report the fraction of excluded trials per method; without this, the central precision/stability comparison is not a fair method-to-method comparison.
  3. [III-D] There is a train/deployment observation mismatch for the lower-body policy that is not analyzed. At deployment, the lower-body RL policy receives the original motion pair (mt, zt), while the reference actually tracked by the upper-body PD controller is the interpolated motion m′_t = linear_interpolate(m_{t−1}, m_{t+Δtt}). During training of the lower-body policy, only original motion pairs are used, so at test time the policy's observation does not describe the motion that is actually being executed. The authors do not quantify the effect of this distribution shift. Either the observation should be updated to include m′_t and its latent encoding, or the lower-body policy should be trained with interpolated references; otherwise the stability and precision guarantees under TOP rest on an unmodeled mismatch.
  4. [IV-C, Table III] The claims about time efficiency are not supported with statistical detail. TOP increases average time cost from 15.0 s to 40.5 s relative to the Ours w/o TOP Δt = 0.01s ablation, which is 2.7× the nominal 15 s motion duration. This trade-off may be acceptable, but every numerical comparison in Table III and Table IV is a single point estimate with no error bars, confidence intervals, or per-clip distributions. The reader cannot assess whether the reported differences (e.g., 95.30% vs. 92.41% success, E_eepe 0.0270 vs. 0.0362) are significant. At minimum the authors should report standard deviations or per-motion distributions, and ideally metrics over the full distribution including failed trials.
minor comments (5)
  1. [IV-C, Table III] Table III has formatting issues in the header (e.g., Eupper jpe, Eupper eepe) and the ablation label 'Ours w/o acting chunking' should be 'Ours w/o action chunking' to match the terminology used in Section III-D.
  2. [III-D, Eq. (6)] Equation (6), Lsup = −Σ_i Δtlog_i πθ(Δt∗_i | s_i), appears to contain a notational error: the supervised loss should be a negative log-likelihood, presumably Lsup = −Σ_i log πθ(Δt∗_i | s_i); as written the expression is not a valid loss.
  3. [Appendix C] Appendix C contains a duplicated word in the sentence 'the humanoid robot executes upper-body tasks and omnidirectional walking' (the text reads 'performs performs'); this should be corrected.
  4. [III-D, Appendix F] The action-chunking weighting formula in Appendix F uses N both as the horizon length and in the normalization sum with i = 0,...,N, while Section III-D defines N as the horizon step and uses exponential weighting w_i = exp(−k·i); the indexing and whether w0 corresponds to the oldest or newest action should be clarified.
  5. [IV, Fig. 5] Figure 5 plots only Exbody and Ours w/o TOP (Δt = 0.01s) while the text compares several baselines; the caption should state why OmniH2O, Mobile-Television, and NMPC+WBC are omitted and whether the omitted baselines show qualitatively similar behavior.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the reported successes are empirical outcomes from held-out motion data and real-robot ablations, not consequences of the training rewards or of self-citations.

full rationale

TOP's central claim, that adaptive timestamp optimization improves standing-manipulation accuracy and stability, is tested against held-out motion dataset T, fixed-timestep ablations, whole-body RL baselines, and real-robot trials (Table III). The RL reward in Table I does encode balance and small-delta-t preferences, but the measured success rate, tracking errors, and time costs are external metrics evaluated on 10,000+ clips and on hardware, so the result is not forced by reward construction. The supervised term Lsup uses the original motion-clip timestamp only as an initialization guess, with independent RL refinement, and the paper explicitly reports worse time efficiency than baselines, showing the outcome is not a fitted restatement of the reward. Self-citations such as [7] provide a baseline and a curriculum heuristic, but the key claims are supported by internal ablations (w/o TOP, w/o motion prior, w/o action chunking) and by comparison with externally cited methods, so no load-bearing argument reduces to a self-citation. One limitation passage should be flagged as a correctness risk rather than circularity: Sec. III-D asserts that linear interpolation of dataset motions "will not violate the constraints" because the dataset satisfies them, while Appendix C concedes the VAE is purely kinematic and relies on the downstream RL policy to compensate dynamic variations. That is an unverified feasibility assumption, but it is not a claim that is defined in terms of its own output, so it does not raise the circularity score.

Assumptions & free parameters 7 free parameters · 6 assumptions · 0 invented entities

The central claim depends on a number of hand-chosen reward weights, hyperparameters (horizon, decay, latent size, delta-t bounds), and domain assumptions about the feasibility of interpolated motion and sim-to-real transfer. None of these are derived; they are empirical design choices. However, the paper does not introduce new physical entities, and its main mechanism (adaptive time scaling) is a control policy rather than a fitted law.

free parameters (7)
  • TOP reward weights (Table I) = 2.5, -1.0, -5.0, 5.0, -0.1, 0.1
    Hand-tuned weights shaping balance, precision, and small-delta-t objectives; directly determine TOP's behavior.
  • Lower-body RL reward weights (Table II) = 3.0, -0.8, -0.1, -1.5, 0.2, 1.0, 0.5, 0.2, -0.2, -0.2, -5e-6, -5e-4, -1e-7
    Hand-designed reward shaping terms for the balance policy; stable results depend on these choices.
  • Loss weights lambda_sup, lambda_RL = 0.1, 0.5
    Weights in Eq. 8 balancing supervised and RL losses for TOP; chosen without a reported tuning study.
  • Delta-t bounds = [0.01 s, 0.1 s]
    Clipping range for the TOP output; restricts the time scaling to this window.
  • Action chunking horizon N and decay k = N=10, k=0.5
    Horizon 'obtained based on our testing'; k selected from a sensitivity analysis to maximize success rate.
  • VAE latent dim and window size = d_z=64, W=30
    Chosen from the sensitivity analysis in Table IX to minimize reconstruction loss; these are the representation's capacity parameters.
  • ZML stability threshold d = 0.32 m
    Ad hoc threshold in Appendix H used to classify the robot as stable in the ZMP/ZML analysis.
assumptions (6)
  • domain assumption Linear interpolation between motion clips preserves kinematic and dynamic feasibility (Section III-D).
    Asserted without proof: 'after linear interpolation, m′_t will not violate the constraints'. The TOP replaces motion frames with interpolated frames, so this is load-bearing.
  • domain assumption The upper-body PD controller tracks reference motions with high precision (Section III-C).
    The decoupled design assumes the PD controller can execute the reference trajectory well enough that precision is not the bottleneck.
  • domain assumption The GRAB human motion dataset, retargeted to the robot, provides a valid distribution of feasible upper-body motions (Section IV-A).
    Assumes retargeting via [35] yields robot-feasible motions; no feasibility check is reported.
  • domain assumption The VAE latent representation captures enough information for the balance policy and TOP (Section III-B, III-D).
    The method depends on z_t conveying past and future motion context; supported only by reconstruction loss and one interpolation figure.
  • domain assumption RL policies trained with domain randomization transfer from simulation to the real robot (Section IV-A, Appendix VI-C).
    Sim-to-real transfer is assumed based on domain randomization; real-world results are used as evidence but without baseline comparisons.
  • standard math Momentum equations in Appendix I (Eq. 17-19) adequately describe upper-body disturbance on the base.
    Newton-Euler relations, but simplified: ignores actuation torques, joint coupling, and contact constraints; used as motivation rather than a rigorous proof.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TOP: Time Optimization Policy for Stable and Accurate Standing Manipulation with Humanoid Robots." pith.science (2026). https://pith.science/paper/YCEHSWD4

@misc{pith2026250800355,
  author       = {Pith},
  title        = {Pith review of: TOP: Time Optimization Policy for Stable and Accurate Standing Manipulation with Humanoid Robots},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YCEHSWD4}},
  note         = {Machine review of arXiv:2508.00355}
}
read the original abstract

Humanoid robots have the potential capability to perform a diverse range of manipulation tasks, but this is based on a robust and precise standing controller. Existing methods are either ill-suited to precisely control high-dimensional upper-body joints, or difficult to ensure both robustness and accuracy, especially when upper-body motions are fast. This paper proposes a novel time optimization policy (TOP), to train a standing manipulation control model that ensures balance, precision, and time efficiency simultaneously, with the idea of adjusting the time trajectory of upper-body motions but not only strengthening the disturbance resistance of the lower-body. Our approach consists of three parts. Firstly, we utilize motion prior to represent upper-body motions to enhance the coordination ability between the upper and lower-body by training a variational autoencoder (VAE). Then we decouple the whole-body control into an upper-body PD controller for precision and a lower-body RL controller to enhance robust stability. Finally, we train TOP method in conjunction with the decoupled controller and VAE to reduce the balance burden resulting from fast upper-body motions that would destabilize the robot and exceed the capabilities of the lower-body RL policy. The effectiveness of the proposed approach is evaluated via both simulation and real world experiments, which demonstrate the superiority on standing manipulation tasks stably and accurately. The project page can be found at https://anonymous.4open.science/w/top-258F/.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

38 extracted references · 24 canonical work pages

  1. [1]

    Humanoid locomotion and manipulation: Current progress and challenges in control, planning, and learning

    Zhaoyuan Gu et al. “Humanoid locomotion and manipulation: Current progress and challenges in control, planning, and learning”. In: arXiv preprint arXiv:2501.02116 (2025)

    arXiv 2025

  2. [2]

    HOMIE: Humanoid Loco-Manipulation with Isomorphic Exoskeleton Cockpit

    Qingwei Ben et al. “HOMIE: Humanoid Loco-Manipulation with Isomorphic Exoskeleton Cockpit”. In: arXiv preprint arXiv:2502.13013 (2025). ��� ��� ��� ��� ��� ��� ��� Fig. 8: Manipulation tasks. (a) grab the cup and put it onto cups, (b) take and deliver a bottle of water, (c) dance with arms but slow down the original motions while the lower body maintains ...

    arXiv 2025

  3. [3]

    Online non-linear centroidal mpc for humanoid robot locomotion with step adjustment

    Giulio Romualdi et al. “Online non-linear centroidal mpc for humanoid robot locomotion with step adjustment”. In: 2022 International Conference on Robotics and Automation (ICRA). IEEE. 2022, pp. 10412–10419

    work page 2022

  4. [4]

    Vmp: Versatile motion priors for robustly tracking motion on physical characters

    Agon Serifi et al. “Vmp: Versatile motion priors for robustly tracking motion on physical characters”. In: Computer Graph- ics Forum. V ol. 43. 8. Wiley Online Library. 2024, e15175

    work page 2024

  5. [5]

    Exbody2: Advanced expressive humanoid whole-body control

    Mazeyu Ji et al. “Exbody2: Advanced expressive humanoid whole-body control”. In: arXiv preprint arXiv:2412.13196 (2024)

    arXiv 2024

  6. [6]

    Deep imitation learning for humanoid loco- manipulation through human teleoperation

    Mingyo Seo et al. “Deep imitation learning for humanoid loco- manipulation through human teleoperation”. In: 2023 IEEE- RAS 22nd International Conference on Humanoid Robots (Humanoids). IEEE. 2023, pp. 1–8

    work page 2023

  7. [7]

    Mobile-television: Predictive motion pri- ors for humanoid whole-body control

    Chenhao Lu et al. “Mobile-television: Predictive motion pri- ors for humanoid whole-body control”. In: arXiv preprint arXiv:2412.07773 (2024)

    arXiv 2024

  8. [8]

    Real-world humanoid locomotion with reinforcement learning

    Ilija Radosavovic et al. “Real-world humanoid locomotion with reinforcement learning”. In: Science Robotics 9.89 (2024), eadi9579

    work page 2024

Show all 38 references
  1. [9]

    HumanPlus: Humanoid Shadowing and Imitation from Humans

    Zipeng Fu et al. “HumanPlus: Humanoid Shadowing and Imitation from Humans”. In: arXiv preprint arXiv:2406.10454 (2024)

    2024 arXiv

  2. [10]

    Visual whole-body control for legged loco-manipulation

    Minghuan Liu et al. “Visual whole-body control for legged loco-manipulation”. In: arXiv preprint arXiv:2403.16967 (2024)

    2024 arXiv

  3. [11]

    Rdt-1b: a diffusion foundation model for bimanual manipulation

    Songming Liu et al. “Rdt-1b: a diffusion foundation model for bimanual manipulation”. In: arXiv preprint arXiv:2410.07864 (2024)

    2024 arXiv

  4. [12]

    Learning fine-grained bimanual ma- nipulation with low-cost hardware

    Tony Z Zhao et al. “Learning fine-grained bimanual ma- nipulation with low-cost hardware”. In: arXiv preprint arXiv:2304.13705 (2023)

    2023 arXiv

  5. [13]

    Bipedal oriented whole body master- slave system for dynamic secured locomotion with lip safety constraints

    Yasuhiro Ishiguro et al. “Bipedal oriented whole body master- slave system for dynamic secured locomotion with lip safety constraints”. In: 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) . IEEE. 2017, pp. 376– 382

    2017

  6. [14]

    A multimode teleoperation framework for humanoid loco-manipulation: An application for the icub robot

    Luigi Penco et al. “A multimode teleoperation framework for humanoid loco-manipulation: An application for the icub robot”. In: IEEE Robotics & Automation Magazine 26.4 (2019), pp. 73–82

    2019

  7. [15]

    Perpetual humanoid control for real-time simulated avatars

    Zhengyi Luo et al. “Perpetual humanoid control for real-time simulated avatars”. In: Proceedings of the IEEE/CVF Inter- national Conference on Computer Vision . 2023, pp. 10895– 10904

    2023

  8. [16]

    Hierarchical World Models as Vi- sual Whole-Body Humanoid Controllers

    Nicklas Hansen et al. “Hierarchical World Models as Vi- sual Whole-Body Humanoid Controllers”. In: arXiv preprint arXiv:2405.18418 (2024)

    2024 arXiv

  9. [17]

    Whole-body humanoid robot locomotion with human reference

    Qiang Zhang et al. “Whole-body humanoid robot locomotion with human reference”. In: 2024 IEEE/RSJ International Con- ference on Intelligent Robots and Systems (IROS). IEEE. 2024, pp. 11225–11231

    2024

  10. [18]

    OmniH2O: Universal and Dexterous Human- to-Humanoid Whole-Body Teleoperation and Learning

    Tairan He et al. “OmniH2O: Universal and Dexterous Human- to-Humanoid Whole-Body Teleoperation and Learning”. In: arXiv preprint arXiv:2406.08858 (2024)

    2024 arXiv

  11. [19]

    Expressive whole-body control for hu- manoid robots

    Xuxin Cheng et al. “Expressive whole-body control for hu- manoid robots”. In: arXiv preprint arXiv:2402.16796 (2024)

    2024 arXiv

  12. [20]

    Deepphase: Periodic autoencoders for learning motion phase manifolds

    Sebastian Starke, Ian Mason, and Taku Komura. “Deepphase: Periodic autoencoders for learning motion phase manifolds”. In: ACM Transactions on Graphics (TOG) 41.4 (2022), pp. 1– 13

    2022

  13. [21]

    Ase: Large-scale reusable adversarial skill embeddings for physically simulated characters

    Xue Bin Peng et al. “Ase: Large-scale reusable adversarial skill embeddings for physically simulated characters”. In: ACM Transactions On Graphics (TOG) 41.4 (2022), pp. 1–17

    2022

  14. [22]

    Amp: Adversarial motion priors for styl- ized physics-based character control

    Xue Bin Peng et al. “Amp: Adversarial motion priors for styl- ized physics-based character control”. In: ACM Transactions on Graphics (ToG) 40.4 (2021), pp. 1–20

    2021

  15. [23]

    Synthesizing physical character- scene interactions

    Mohamed Hassan et al. “Synthesizing physical character- scene interactions”. In: ACM SIGGRAPH 2023 Conference Proceedings. 2023, pp. 1–9

    2023

  16. [24]

    Humanmimic: Learning natural locomotion and transitions for humanoid robot via wasserstein adversar- ial imitation

    Annan Tang et al. “Humanmimic: Learning natural locomotion and transitions for humanoid robot via wasserstein adversar- ial imitation”. In: 2024 IEEE International Conference on Robotics and Automation (ICRA) . IEEE. 2024, pp. 13107– 13114

    2024

  17. [25]

    Robust motion in-betweening

    F ´elix G Harvey et al. “Robust motion in-betweening”. In:ACM Transactions on Graphics (TOG) 39.4 (2020), pp. 60–1

    2020

  18. [26]

    Emdm: Efficient motion diffusion model for fast and high-quality motion generation

    Wenyang Zhou et al. “Emdm: Efficient motion diffusion model for fast and high-quality motion generation”. In: European Conference on Computer Vision . Springer. 2024, pp. 18–38

    2024

  19. [27]

    Physics-based character controllers using conditional vaes

    Jungdam Won, Deepak Gopinath, and Jessica Hodgins. “Physics-based character controllers using conditional vaes”. In: ACM Transactions on Graphics (TOG) 41.4 (2022), pp. 1– 12

    2022

  20. [28]

    Human motion diffusion model

    Guy Tevet et al. “Human motion diffusion model”. In: arXiv preprint arXiv:2209.14916 (2022)

    2022 arXiv

  21. [29]

    GENMO: A GENeralist Model for Human MOtion

    Jiefeng Li et al. “GENMO: A GENeralist Model for Human MOtion”. In: arXiv preprint arXiv:2505.01425 (2025)

    2025 arXiv

  22. [30]

    beta-vae: Learning basic visual concepts with a constrained variational framework

    Irina Higgins et al. “beta-vae: Learning basic visual concepts with a constrained variational framework.” In: ICLR (Poster) 3 (2017)

    2017

  23. [31]

    Hybrid internal model: Learning agile legged locomotion with simulated robot response

    Junfeng Long et al. “Hybrid internal model: Learning agile legged locomotion with simulated robot response”. In: arXiv preprint arXiv:2312.11460 (2023)

    2023 arXiv

  24. [32]

    Action chunking as policy compression

    Lucy Lai, Ann Zixiang Huang, and Samuel J Gershman. “Action chunking as policy compression”. In: (2022)

    2022

  25. [33]

    Supervised reinforcement learning with recur- rent neural network for dynamic treatment recommendation

    Lu Wang et al. “Supervised reinforcement learning with recur- rent neural network for dynamic treatment recommendation”. In: Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining . 2018, pp. 2447–2456

    2018

  26. [34]

    GRAB: A dataset of whole-body human grasping of objects

    Omid Taheri et al. “GRAB: A dataset of whole-body human grasping of objects”. In: Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part IV 16 . Springer. 2020, pp. 581–600

    2020

  27. [35]

    Kinematic motion retargeting via neu- ral latent optimization for learning sign language

    Haodong Zhang et al. “Kinematic motion retargeting via neu- ral latent optimization for learning sign language”. In: IEEE Robotics and Automation Letters 7.2 (2022), pp. 4582–4589

    2022

  28. [36]

    Perceptive locomotion through nonlinear model-predictive control

    Ruben Grandia et al. “Perceptive locomotion through nonlinear model-predictive control”. In: IEEE Transactions on Robotics 39.5 (2023), pp. 3402–3421

    2023

  29. [37]

    A computational multicriteria optimiza- tion approach to controller design for physical human-robot interaction

    Yusuf Aydin et al. “A computational multicriteria optimiza- tion approach to controller design for physical human-robot interaction”. In: IEEE Transactions on Robotics 36.6 (2020), pp. 1791–1804

    2020

  30. [38]

    Zero moment line—universal stability parameter for multi-contact systems in three dimen- sions

    Tilen Brecelj and Tadej Petri ˇc. “Zero moment line—universal stability parameter for multi-contact systems in three dimen- sions”. In: Sensors 22.15 (2022), p. 5656. VI. APPENDIX A. Motion Alignment And Coupling Effects In principle, upper- and lower-body motions are inherent...

    2022

Pith tools

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