EMP: Executable Motion Prior for Humanoid Robot Standing Upper-body Motion Imitation

Haocheng Xu; Haodong Zhang; Rong Xiong; zhenghan Chen

REVIEW 3 major objections 4 minor 38 references

EMP: Executable Motion Prior for Humanoid Robot Standing Upper-body Motion Imitation

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A state-conditioned filter turns unexecutable upper-body motion targets into stable humanoid imitation.

desk verdict The EMP module is a sensible new combination, but the deployment section appears to validate a baseline and the unvalidated world model undermines the training pipeline. read the letter →

arxiv 2507.15649 v1 pith:4JAB3J4B submitted 2025-07-21 cs.RO

Haocheng Xu , Haodong Zhang , Zhenghan Chen , Rong Xiong This is my paper

classification cs.RO

keywords humanoidrobotsupper-bodymotionimitationreinforcementlearningexecutablepriorvariationalautoencoderworldmodelretargetingsim-to-realtransfer

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 tries to establish that a humanoid robot standing in place can imitate human upper-body motions—arm, torso, and waist movements—without falling, by separating the problem into a lower-body balance controller and a learnable filter for the upper-body targets. The filter, called the Executable Motion Prior (EMP), reads the robot's current state and rewrites each incoming motion target so that the target stays close to the original motion but is actually within the robot's balancing ability. The authors train an RL policy to control only the legs while the filtered arm, torso, and waist targets are executed directly, and they show in simulation that the full system reaches a 98.1 percent success rate on the primary robot, up from 97.0 percent for the same policy without the filter, with better base velocity, acceleration, and orientation stability. They also deploy the framework on two physical humanoid robots. If the claim holds, it gives a practical route from human motion capture to stable upper-body manipulation on full-sized humanoids.

What carries the argument

The load-bearing object is the Executable Motion Prior (EMP), a conditional variational autoencoder. It has a state encoder and a target encoder that map $s_t$ and $g_t$ into latent variables $z_1$ and $z_2$, a fusion network that combines them into a single latent vector $z$, and a decoder that outputs the revised target $\hat{g}_t$. Because the simulator does not provide gradients from the robot state, the authors introduce a world model $f_w(s_t,a_t)$ that predicts the next state; EMP training backpropagates the orientation and centroid losses through this model, while the reconstruction, collision, smoothness, and regularization losses keep the revision faithful and feasible. The same retargeting and RL components supply the data and the lower-body policy that the EMP sits in front of.

What would settle it

Measure the next-state predictor's one-step and multi-step error against the simulator and against the real robot on the test motions. If its predicted orientation or centroid errors are large or biased under the motions the EMP actually generates, then the orientation and centroid losses are being optimized on the wrong quantity, and the 98.1 percent success rate should not survive a policy trained with a more accurate model.

Watch

Extended reading notes

Core claim

The central claim is that target motions for a standing humanoid can be made executable by a learned state-conditioned filter. The EMP network computes $\hat{g}_t = f_\theta(s_t, g_t)$: it encodes the robot's proprioceptive state $s_t$ and the desired upper-body target $g_t$ into a latent code and decodes a revised target $\hat{g}_t$. It is trained so that the revised target minimizes changes from the original (reconstruction loss $L_{\text{rec}}$) while simultaneously driving the world-model-predicted next state toward an upright base ($L_{\text{ori}}$), a centroid over the foot support area ($L_{\text{cen}}$), fewer self-collisions ($L_{\text{col}}$), and smooth joint motion ($L_{\text{smo}}$), with the world model $\hat{s}_{t+1} = f_w(s_t, a_t)$ providing gradients from the next state. The authors report that this raises success rate on the primary robot from 97.0 percent for the decoupled policy alone to 98.1 percent, improves base velocity, acceleration, and orientation rewards, and slightly increases joint-position error. Removing the orientation or centroid losses collapses success to below 11 percent. On a second robot the gain is larger, from 90.0 percent to 97.8 percent, and the same system is shown working in real-world tests.

Load-bearing premise

The procedure assumes the learned predictor of the robot's next state—especially its tilt and center of mass—is accurate enough that the stability improvements computed through it are real improvements on the actual robot, yet the paper reports no direct check of that predictor's error.

Editorial extensions

If this is right

A standing humanoid can track retargeted human arm, torso, and waist motions through a decoupled design: an RL policy stabilizes the legs while a state-conditioned filter governs what the upper body is asked to do.
Filtering targets before they reach the controller is enough to recover most of the stability lost to unexecutable motions, with only a small increase in upper-body tracking error.
The orientation and centroid losses are not optional: removing either one drops simulated success to single digits or low tens of percent on the primary robot.
The framework transfers: the same recipe improves success on a second humanoid from 90.0 percent to 97.8 percent, and the deployed system imitates motions on physical robots.

Reading between the lines

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

The reported numbers suggest that the EMP's value scales with how close the baseline controller is to its balance limits, so a fair cross-platform comparison would test it on a deliberately weak lower-body policy.
The paper records no direct validation of the world model's prediction error, even though every EMP gradient passes through it; a natural test is to compare one-step and multi-step world-model predictions against the simulator and hardware and to retrain with an uncertainty-weighted loss.
The 'EMP when Danger' variant shows the filter can be run selectively based on the latent regularity score, suggesting the same mechanism could serve as a safety filter in teleoperation, intervening only when motion targets leave the executable distribution.
The paper's stated limitations—no whole-body imitation and a visible gap between retargeted and source motions due to joint limits—imply that extending the prior to leg and foot targets is the natural next step, where balance is even more tightly coupled to the motion being imitated.

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.

Desk Editor's Note

read the letter

The short version: the central idea—a state-conditioned VAE that rewrites upper-body targets to keep a standing humanoid stable—is a reasonable integration of existing pieces, and the simulated results are encouraging. But the real-world section says the deployed system was PMP (baseline iv), not EMP, and the world model that the entire EMP training loop depends on is never validated. Both need fixing before I'd believe the headline claim.

What's actually new: the specific combination of VQ-VAE retargeting, a lower-body RL balance policy, and the EMP filter trained through a learned world model. The ablation study is the strongest part—removing smoothness, orientation, or centroid losses collapses success rates from 98% to 27%, 2.6%, and 10.7%, which shows those losses are load-bearing. The retargeting network follows Zhang et al. [17], but the EMP itself is a genuine new module.

Now the soft spots. Section VI-C states: "The algorithm we deployed on the real-world system is baseline iv)." In the baseline list, iv) is PMP, not the proposed EMP (v) or EMP-when-Danger (vi). If that is not a typo, the hardware demonstration validates a baseline, not the paper's contribution. The authors must clarify this. Second, the world model f_w is trained with L_pre, but no prediction error is reported anywhere, despite Algorithm 1 using it for every EMP gradient step. If f_w is inaccurate or biased, the orientation and centroid losses optimize the wrong quantity. The stress-test's worry that f_w's input st contains the original target rather than the modified ĝ_t is less damning—the action at is sampled from π(st, ĝ_t), so gradients can flow through the action—but the paper does not explain this, and the architecture is underdescribed. Third, the metrics in Table IV overlap heavily with the EMP training losses, so part of the measured gain is expected by construction. Finally, the simulation results come from single runs without error bars; the 98.1% vs 97.0% difference could be noise, though the ablation gaps are large enough to be convincing.

Who this paper is for: people working on humanoid upper-body manipulation, imitation, or sim-to-real control. It deserves a serious peer review—the idea is useful and the ablations are informative—but the authors should be asked to release code and data, report multi-seed statistics, validate the world model, and fix the deployment description. If the deployment text is a typo, it's a small fix; if not, the real-world claim collapses.

Referee Report

3 major / 4 minor

Summary. The paper proposes EMP, an Executable Motion Prior for humanoid robots imitating human upper-body motions while standing. The framework has three components: a graph-convolutional retargeting network that converts human motion into humanoid upper-body joint targets (Section III); a goal-conditioned RL policy that controls the lower body to maintain balance while tracking those targets (Section IV); and a VAE-based EMP network that adjusts the upper-body target online based on the robot's current state (Section V). A learned world model is used to backpropagate stability losses through a differentiable state transition. The authors report simulation experiments on two humanoid platforms (Tables IV and VI), an ablation study (Table V), and real-world deployment on two robots (Section VI-C and VI-D), claiming that EMP improves standing stability with minimal change to motion amplitude.

Significance. If the central claim holds, the work is a useful practical contribution: it addresses an important gap in upper-body motion imitation for humanoid robots, where directly executing upper-body targets can destabilize the robot. The framework combines retargeting, RL lower-body control, and a learned motion prior in a modular way, and the paper includes domain randomization and deployment on physical robots. The retargeting network and EMP losses are clearly described, and the ablation study in Table V gives some evidence that the proposed losses matter. However, the evidence for the central claim is substantially weakened by three issues: the real-world deployment section states that the deployed algorithm was PMP, not EMP; the world model used for every EMP gradient step is unvalidated and its input omits the modified upper-body target; and the reported evaluation metrics overlap heavily with the training losses, while results come from single runs without variance estimates.

major comments (3)

[Section VI-C] The real-world deployment does not actually test EMP. Section VI-C states: 'The algorithm we deployed on the real-world system is baseline iv).' In Section VI-A, baseline iv) is 'Decoupled Imitation Policy with Predictive Motion Prior (PMP) [10]', not the proposed EMP (baseline v) or its 'EMP when Danger' variant (baseline vi). As written, the hardware demonstrations support PMP, not EMP, so the abstract's claim of 'real-world tests' validating EMP's stability improvement is not supported. This needs to be corrected if it is a typo, or the real-world claims must be re-scoped if it is not.
[Section V-B and Algorithm 1] The world-model gradient path appears to be inconsistent with the EMP mechanism. In Section IV-B the observation state is defined as s_t = [q_t, a_{t-1}, rpy_t, g_t], so s_t contains the original upper-body target g_t. In Algorithm 1, line 12 samples a_t = π(s_t, g_hat_t), but line 13 then evaluates f_w(s_t, a_t) without providing g_hat_t to the world model. The next physical state depends on the modified upper-body target that is actually sent to the robot, yet f_w cannot see it. Consequently, the orientation loss (Eq. 7) and centroid loss (Eq. 9) propagated through f_w may optimize the wrong quantity. Moreover, the paper reports no validation of f_w's prediction error against the simulator or the real robot, even though f_w is used for every EMP gradient step.
[Table IV and Section V-B] The evaluation metrics are closely aligned with the quantities explicitly optimized by EMP, so part of the reported improvement is expected by construction. The orientation loss (Eq. 7), collision loss (Eq. 8), smoothness loss (Eq. 10), and centroid loss (Eq. 9) correspond directly to MBO, MSC, MUS, and success/balance measures in Table IV. For example, EMP improves MBO from 0.6699 to 0.6892 and MSC from 0.3668 to 0.1494, exactly the terms it was trained to minimize. The central stability claim would be much stronger with an independent metric (e.g., zero-moment-point margin, reaction-force margin, or success under external pushes) and with multiple training seeds or confidence intervals, especially because the headline success-rate gain is only 97.0% to 98.1%.

minor comments (4)

[References] Reference [10] is titled 'Mobile-television: Predictive motion priors for humanoid whole-body control'; this appears to be a typo for 'Mobile-TeleVision' and should be corrected.
[Figure 2] The caption says 'Vq-Layer' while Section III-A uses 'vq-codebook layer'; please unify the terminology.
[Tables IV and VI] The tables report metrics to inconsistent precision (e.g., Table IV uses 0.8221 while Table V uses 0.822; Table VI uses 0.840 but 0.861). Standardize the number of decimal places to make comparison easier.
[Section VI-D] The sentence 'The results are shown in Figure 7' refers to simulation results on another platform, but the preceding paragraph discusses deployment on another humanoid platform; clarify whether Figure 7 shows real-world or simulation results.

Circularity Check

1 steps flagged · score 5.0 of 10

EMP's reported stability gains partly coincide with its own training losses, but the central claim retains independent baseline comparisons.

fitted input called prediction [Section V-B Eqs. (7)-(9), Section IV-E, Section VI-A Metrics and Table IV]
"ii) Orientation Loss. The orientation loss Lori promotes the humanoid’s base to stay upright ... Lori = exp(−∥cpgxy t+1∥2 2) − 1 ... iv) Centroid Loss. ... Lcen = min{exp(−7(0.03 − d)), 10} −1 ... We define imitation failed when termination conditions in section IV-E are triggered. ... we reset training process when the projected gravity on x or y axis exceeds 0.7."

EMP is trained by minimizing Eq. (7), a monotone function of the projected gravity vector cpg^xy_{t+1}, and Eq. (9), a monotone function of the centroid-projection distance d. The same physical quantities define the paper's headline stability evaluation: SUC is defined as avoiding the Section IV-E termination condition, which is triggered when 'projected gravity on x or y axis exceeds 0.7', and MBO is the base-orientation reward built from roll, pitch, and projected gravity. Thus the reported SUC and MBO gains of EMP over the decoupled policy (97.0% to 98.1% and 0.6699 to 0.6892) are partly forced by construction: the module is trained to minimize the very functions used to score it.

full rationale

The score is 5 rather than 0 because one component of the central claim, that EMP improves standing stability, is evaluated with metrics that are nearly identical to the EMP training losses. Section V-B's orientation loss (Eq. 7) and centroid loss (Eq. 9) optimize the projected gravity and centroid projection, while Section VI-A's SUC metric is defined by the same projected-gravity termination threshold used in Section IV-E, and MBO is the corresponding orientation reward. Consequently, the SUC and MBO deltas in Table IV are partly a check that the training objective decreased, not an independent prediction. I do not count this as full circularity: the paper compares against external baselines (PMP, whole-body policy, privileged policy), reports independent metrics (MBV, MBA, MUS, MSC, MJP), and the retargeting and RL-policy components are not defined in terms of the EMP output. The self-citations [17] and [27] support only the retargeting network, not the EMP claim, so they are not load-bearing for the central stability result. Separately, the real-world section states that 'the algorithm we deployed on the real-world system is baseline iv)', which is PMP rather than EMP; if accurate, this weakens the hardware evidence but is not a circularity and is not scored here. The unvalidated world model is also a correctness risk rather than a circular step, since gradients through f_w do not reduce to the evaluation metric by definition.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

No new physical entities or conserved quantities are introduced. EMP and the world model are computational components, not entities with independent falsifiable handles outside the paper.

free parameters (5)

EMP loss weights = lambda_rec=20, lambda_ori=10, lambda_col=1, lambda_cen=10, lambda_smo=100, lambda_reg=1
Hand-chosen in Section V-B; not tuned systematically and no sensitivity analysis is given. They set the trade-off between preserving the original motion and satisfying the stability losses.
EMP-when-Danger threshold = 0.04 on latent regloss
Used in the EMP-when-Danger baseline to enable the prior only when the latent code deviates from the prior; chosen post hoc without a search or robustness check.
Centroid support margin = 0.03 m
Constant in centroid loss Eq. 9; hand-set and directly affects what is considered 'inside' the support region.
Collision buffer distance = 0.08 m
Constant in collision loss Eq. 8; hand-set and defines when links are treated as colliding.
Termination criterion for projected gravity = 0.7 on x/y projected gravity
Defines success rate in Section IV-E and all SUC metrics; hand-set, and the EMP orientation loss uses a similar metric, so improvements may be concentrated near this threshold.

assumptions (4)

ad hoc to paper The world model f_w (Eq. 4) accurately predicts the next robot state so that gradients through f_w improve EMP.
Invoked in Algorithm 1 and Section V-B for all EMP gradient updates; no validation of f_w accuracy is reported.
domain assumption Projected gravity, centroid projection, and collision penalties are sufficient proxies for real standing balance.
Used in EMP losses (Section V-B, Eqs. 7-9) and termination conditions (Section IV-E); the paper does not validate these against real-world balance failures.
domain assumption The retargeted GRAB/AMASS motions are a sufficient training distribution for standing upper-body imitation.
Section VI-A trains both RL and EMP on this retargeted dataset; no evaluation on held-out or out-of-distribution movement classes is reported.
domain assumption Isaac Gym simulation with domain randomization transfers to the real robot without additional adaptation.
Section IV-D and Section VI-C assume sim-to-real transfer; real-world results are qualitative, so this transfer is not quantitatively verified.

how reviews work

0 comments

Cite this review

Pith. "Pith review of EMP: Executable Motion Prior for Humanoid Robot Standing Upper-body Motion Imitation." pith.science (2026). https://pith.science/paper/4JAB3J4B

@misc{pith2026250715649,
  author       = {Pith},
  title        = {Pith review of: EMP: Executable Motion Prior for Humanoid Robot Standing Upper-body Motion Imitation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4JAB3J4B}},
  note         = {Machine review of arXiv:2507.15649}
}

read the original abstract

To support humanoid robots in performing manipulation tasks, it is essential to study stable standing while accommodating upper-body motions. However, the limited controllable range of humanoid robots in a standing position affects the stability of the entire body. Thus we introduce a reinforcement learning based framework for humanoid robots to imitate human upper-body motions while maintaining overall stability. Our approach begins with designing a retargeting network that generates a large-scale upper-body motion dataset for training the reinforcement learning (RL) policy, which enables the humanoid robot to track upper-body motion targets, employing domain randomization for enhanced robustness. To avoid exceeding the robot's execution capability and ensure safety and stability, we propose an Executable Motion Prior (EMP) module, which adjusts the input target movements based on the robot's current state. This adjustment improves standing stability while minimizing changes to motion amplitude. We evaluate our framework through simulation and real-world tests, demonstrating its practical applicability.

Figures

Figures reproduced from arXiv: 2507.15649 by the authors.

**Figure 2.** Overview of our framework. Motion Retargeting (section III): We train a graph convolution retargeting network to convert human motions msource to humanoid joint actions mtarget as motion goal for imitation. RL Policy (section IV): We train an upper-body imitation policy for the humanoid to track the upper-body motion goal gt while keeping balance. Executable Motion Prior (section V): We use a VAE-based network to ad… view at source ↗

**Figure 3.** Framework of our EMP System. EMP network gener [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Figures from the paper (4 more)

**Figure 4.** Figure 4: Simulation experiments (left motion: hammer use, right motion: lightbulb screw). The results show that while executing [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

**Figure 5.** Figure 5: The upper motion (lightbulb screw) of whole-body [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

**Figure 6.** Figure 6: Humanoid robot imitating dataset motion. [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]

**Figure 7.** Figure 7: Simulation experiment on another platform [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

38 extracted references · 16 canonical work pages

[17]

Unified cross-structural motion retargeting for humanoid char- acters,

H. Zhang, Z. Chen, H. Xu, L. Hao, X. Wu, S. Xu, R. Xiong, and Y . Wang, “Unified cross-structural motion retargeting for humanoid char- acters,” IEEE Transactions on Visualization and Computer Graphics , pp. 1–14, 2024

work page 2024
[10]

Mobile-television: Predictive motion priors for humanoid whole-body control,

C. Lu, X. Cheng, J. Li, S. Yang, M. Ji, C. Yuan, G. Yang, S. Yi, and X. Wang, “Mobile-television: Predictive motion priors for humanoid whole-body control,” 2025. [Online]. Available: https: //arxiv.org/abs/2412.07773

arXiv 2025
[1]

Bilateral humanoid teleoperation system using whole-body exoskeleton cockpit tablis,

Y . Ishiguro, T. Makabe, Y . Nagamatsu, Y . Kojio, K. Kojima, F. Sugai, Y . Kakiuchi, K. Okada, and M. Inaba, “Bilateral humanoid teleoperation system using whole-body exoskeleton cockpit tablis,” IEEE Robotics and Automation Letters , vol. 5, no. 4, pp. 6419–6426, 2020

work page 2020
[2]

High speed whole body dynamic motion experiment with real time master-slave humanoid robot system,

Y . Ishiguro, K. Kojima, F. Sugai, S. Nozawa, Y . Kakiuchi, K. Okada, and M. Inaba, “High speed whole body dynamic motion experiment with real time master-slave humanoid robot system,” in 2018 IEEE International Conference on Robotics and Automation (ICRA) , 2018, pp. 5835–5841

work page 2018
[3]

Humanoid dynamic synchronization through whole-body bilateral feedback teleoperation,

J. Ramos and S. Kim, “Humanoid dynamic synchronization through whole-body bilateral feedback teleoperation,” IEEE Transactions on Robotics, vol. 34, no. 4, pp. 953–965, 2018

work page 2018
[4]

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

X. B. Peng, Y . Guo, L. Halper, S. Levine, and S. Fidler, “Ase: large- scale reusable adversarial skill embeddings for physically simulated characters,” ACM Transactions on Graphics, vol. 41, no. 4, p. 1–17, July

work page
[5]

Amp: adversarial motion priors for stylized physics-based character control,

X. B. Peng, Z. Ma, P. Abbeel, S. Levine, and A. Kanazawa, “Amp: adversarial motion priors for stylized physics-based character control,” ACM Transactions on Graphics , vol. 40, no. 4, p. 1–20, July 2021. [Online]. Available: http://dx.doi.org/10.1145/3450626.3459670

arXiv 2021
[6]

Blind bipedal stair traversal via sim-to-real reinforcement learning,

J. Siekmann, K. Green, J. Warila, A. Fern, and J. Hurst, “Blind bipedal stair traversal via sim-to-real reinforcement learning,” 2021. [Online]. Available: https://arxiv.org/abs/2105.08328

arXiv 2021

Show all 38 references

[7]

Reinforcement learning for versatile, dynamic, and robust bipedal locomotion control,

Z. Li, X. B. Peng, P. Abbeel, S. Levine, G. Berseth, and K. Sreenath, “Reinforcement learning for versatile, dynamic, and robust bipedal locomotion control,” 2024. [Online]. Available: https: //arxiv.org/abs/2401.16889

2024 arXiv
[8]

Expressive whole-body control for humanoid robots,

X. Cheng, Y . Ji, J. Chen, R. Yang, G. Yang, and X. Wang, “Expressive whole-body control for humanoid robots,” 2024. [Online]. Available: https://arxiv.org/abs/2402.16796

2024 arXiv
[9]

Learning human-to-humanoid real-time whole-body teleoperation,

T. He, Z. Luo, W. Xiao, C. Zhang, K. Kitani, C. Liu, and G. Shi, “Learning human-to-humanoid real-time whole-body teleoperation,”
[11]

Humanplus: Humanoid shadowing and imitation from humans,

Z. Fu, Q. Zhao, Q. Wu, G. Wetzstein, and C. Finn, “Humanplus: Humanoid shadowing and imitation from humans,” in arXiv, 2024

2024
[12]

Physically based motion transformation,

Z. Popovi ´c and A. Witkin, “Physically based motion transformation,” in Proceedings of the 26th Annual Conference on Computer Graphics and Interactive Techniques , ser. SIGGRAPH ’99. USA: ACM Press/Addison-Wesley Publishing Co., 1999, pp. 11–20. [Online]. Available: https://d...

1999
[13]

Correspondence-free online human motion retargeting,

R. Rekik, M. Marsot, A.-H. Olivier, J.-S. Franco, and S. Wuhrer, “Correspondence-free online human motion retargeting,” 2024. [Online]. Available: https://arxiv.org/abs/2302.00556

2024 arXiv
[14]

A deep learning framework for character motion synthesis and editing,

D. Holden, J. Saito, and T. Komura, “A deep learning framework for character motion synthesis and editing,” ACM Trans. Graph. , vol. 35, no. 4, jul 2016. [Online]. Available: https://doi.org/10.1145/2897824. 2925975

2016 doi
[15]

A variational u-net for motion retargeting,

H. Jang, B. Kwon, M. Yu, S. U. Kim, and J. Kim, “A variational u-net for motion retargeting,” in SIGGRAPH Asia 2018 Posters , ser. SA ’18. New York, NY , USA: Association for Computing Machinery, 2018. [Online]. Available: https://doi.org/10.1145/3283289.3283316

2018
[16]

Skeleton-aware networks for deep motion retargeting,

K. Aberman, P. Li, D. Lischinski, O. Sorkine-Hornung, D. Cohen-Or, and B. Chen, “Skeleton-aware networks for deep motion retargeting,” ACM Transactions on Graphics , vol. 39, no. 4, Aug. 2020. [Online]. Available: http://dx.doi.org/10.1145/3386569.3392462

2020
[18]

Transfer learning of shared latent spaces between robots with similar kinematic structure,

B. Delhaisse, D. Esteban, L. Rozo, and D. Caldwell, “Transfer learning of shared latent spaces between robots with similar kinematic structure,” in 2017 International Joint Conference on Neural Networks (IJCNN) , 2017, pp. 4142–4149

2017
[19]

Motion retargeting for humanoid robots based on simultaneous morphing parameter identification and motion optimization,

K. Ayusawa and E. Yoshida, “Motion retargeting for humanoid robots based on simultaneous morphing parameter identification and motion optimization,” IEEE Transactions on Robotics , vol. 33, no. 6, pp. 1343– 1357, 2017

2017
[20]

Perpetual humanoid control for real-time simulated avatars,

Z. Luo, J. Cao, A. Winkler, K. Kitani, and W. Xu, “Perpetual humanoid control for real-time simulated avatars,” 2023. [Online]. Available: https://arxiv.org/abs/2305.06456

2023 arXiv
[21]

Residual force control for agile human behavior imitation and extended motion synthesis,

Y . Yuan and K. M. Kitani, “Residual force control for agile human behavior imitation and extended motion synthesis,” in Proceedings of the 34th International Conference on Neural Information Processing Systems, ser. NIPS ’20. Red Hook, NY , USA: Curran Associates Inc., 2020

2020
[22]

Learning humanoid locomotion with transformers,

I. Radosavovic, T. Xiao, B. Zhang, T. Darrell, J. Malik, and K. Sreenath, “Learning humanoid locomotion with transformers,” arXiv:2303.03381, 2023

2023 arXiv
[23]

Teleoperation of humanoid robots: A survey,

K. Darvish, L. Penco, J. Ramos, R. Cisneros, J. Pratt, E. Yoshida, S. Ivaldi, and D. Pucci, “Teleoperation of humanoid robots: A survey,” IEEE Transactions on Robotics , vol. 39, no. 3, pp. 1706–1727, 2023

2023
[24]

Slomo: A general system for legged robot motion imitation from casual videos,

J. Z. Zhang, S. Yang, G. Yang, A. L. Bishop, D. Ramanan, and Z. Manchester, “Slomo: A general system for legged robot motion imitation from casual videos,” 2023. [Online]. Available: https://arxiv.org/abs/2304.14389

2023 arXiv
[25]

Dynamic locomotion synchronization of bipedal robot and human operator via bilateral feedback teleoperation,

J. Ramos and S. Kim, “Dynamic locomotion synchronization of bipedal robot and human operator via bilateral feedback teleoperation,” Science Robotics, vol. 4, no. 35, p. eaav4282, 2019. [Online]. Available: https://www.science.org/doi/abs/10.1126/scirobotics.aav4282

2019 doi
[26]

Neural discrete representation learning,

A. van den Oord, O. Vinyals, and K. Kavukcuoglu, “Neural discrete representation learning,” 2018. [Online]. Available: https: //arxiv.org/abs/1711.00937

2018 arXiv
[27]

Kinematic motion retargeting via neural latent optimization for learning sign language,

H. Zhang, W. Li, J. Liu, Z. Chen, Y . Cui, Y . Wang, and R. Xiong, “Kinematic motion retargeting via neural latent optimization for learning sign language,” 2022. [Online]. Available: https://arxiv.org/abs/ 2103.08882

2022 arXiv
[28]

Auto-encoding variational bayes,

D. P. Kingma and M. Welling, “Auto-encoding variational bayes,”
[29]

Controlvae: Controllable variational autoencoder,

H. Shao, S. Yao, D. Sun, A. Zhang, S. Liu, D. Liu, J. Wang, and T. Abdelzaher, “Controlvae: Controllable variational autoencoder,”
[30]

Isaac gym: High performance gpu-based physics simulation for robot learning,

V . Makoviychuk, L. Wawrzyniak, Y . Guo, M. Lu, K. Storey, M. Macklin, D. Hoeller, N. Rudin, A. Allshire, A. Handa, and G. State, “Isaac gym: High performance gpu-based physics simulation for robot learning,” 2021. [Online]. Available: https://arxiv.org/abs/2108.10470

2021 arXiv
[31]

Available: https://arxiv.org/abs/1312.6114

[Online]. Available: https://arxiv.org/abs/1312.6114

arXiv
[32]

AMASS: Archive of motion capture as surface shapes,

N. Mahmood, N. Ghorbani, N. F. Troje, G. Pons-Moll, and M. J. Black, “AMASS: Archive of motion capture as surface shapes,” inInternational Conference on Computer Vision , Oct. 2019, pp. 5442–5451

2019
[33]

Adam: A method for stochastic optimization,

D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,”
[35]

Taheri, N

O. Taheri, N. Ghorbani, M. J. Black, and D. Tzionas, GRAB: A Dataset of Whole-Body Human Grasping of Objects . Springer International Publishing, 2020, pp. 581–600. [Online]. Available: http://dx.doi.org/10.1007/978-3-030-58548-8_34

2020 doi
[2017]

Available: https://arxiv.org/abs/1412.6980

[Online]. Available: https://arxiv.org/abs/1412.6980

arXiv
[2020]

Available: https://arxiv.org/abs/2004.05988

[Online]. Available: https://arxiv.org/abs/2004.05988

2004 arXiv
[2022]

Available: http://dx.doi.org/10.1145/3528223.3530110

[Online]. Available: http://dx.doi.org/10.1145/3528223.3530110
[2024]

Available: https://arxiv.org/abs/2403.04436

[Online]. Available: https://arxiv.org/abs/2403.04436

arXiv

Pith tools

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

[1] [17]

Unified cross-structural motion retargeting for humanoid char- acters,

H. Zhang, Z. Chen, H. Xu, L. Hao, X. Wu, S. Xu, R. Xiong, and Y . Wang, “Unified cross-structural motion retargeting for humanoid char- acters,” IEEE Transactions on Visualization and Computer Graphics , pp. 1–14, 2024

work page 2024

[2] [10]

Mobile-television: Predictive motion priors for humanoid whole-body control,

C. Lu, X. Cheng, J. Li, S. Yang, M. Ji, C. Yuan, G. Yang, S. Yi, and X. Wang, “Mobile-television: Predictive motion priors for humanoid whole-body control,” 2025. [Online]. Available: https: //arxiv.org/abs/2412.07773

arXiv 2025

[3] [1]

Bilateral humanoid teleoperation system using whole-body exoskeleton cockpit tablis,

Y . Ishiguro, T. Makabe, Y . Nagamatsu, Y . Kojio, K. Kojima, F. Sugai, Y . Kakiuchi, K. Okada, and M. Inaba, “Bilateral humanoid teleoperation system using whole-body exoskeleton cockpit tablis,” IEEE Robotics and Automation Letters , vol. 5, no. 4, pp. 6419–6426, 2020

work page 2020

[4] [2]

High speed whole body dynamic motion experiment with real time master-slave humanoid robot system,

Y . Ishiguro, K. Kojima, F. Sugai, S. Nozawa, Y . Kakiuchi, K. Okada, and M. Inaba, “High speed whole body dynamic motion experiment with real time master-slave humanoid robot system,” in 2018 IEEE International Conference on Robotics and Automation (ICRA) , 2018, pp. 5835–5841

work page 2018

[5] [3]

Humanoid dynamic synchronization through whole-body bilateral feedback teleoperation,

J. Ramos and S. Kim, “Humanoid dynamic synchronization through whole-body bilateral feedback teleoperation,” IEEE Transactions on Robotics, vol. 34, no. 4, pp. 953–965, 2018

work page 2018

[6] [4]

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

X. B. Peng, Y . Guo, L. Halper, S. Levine, and S. Fidler, “Ase: large- scale reusable adversarial skill embeddings for physically simulated characters,” ACM Transactions on Graphics, vol. 41, no. 4, p. 1–17, July

work page

[7] [5]

Amp: adversarial motion priors for stylized physics-based character control,

X. B. Peng, Z. Ma, P. Abbeel, S. Levine, and A. Kanazawa, “Amp: adversarial motion priors for stylized physics-based character control,” ACM Transactions on Graphics , vol. 40, no. 4, p. 1–20, July 2021. [Online]. Available: http://dx.doi.org/10.1145/3450626.3459670

arXiv 2021

[8] [6]

Blind bipedal stair traversal via sim-to-real reinforcement learning,

J. Siekmann, K. Green, J. Warila, A. Fern, and J. Hurst, “Blind bipedal stair traversal via sim-to-real reinforcement learning,” 2021. [Online]. Available: https://arxiv.org/abs/2105.08328

arXiv 2021