Pith. sign in

REVIEW 5 major objections 6 minor 37 references

Generating Physically Realistic and Directable Human Motions from Multi-Modal Inputs

T0 review · 5 major / 6 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read One physics-based controller, trained on selectively masked motion demonstrations, can catch up to out-of-sync commands, combine body segments across motions, and complete sparse multimodal directives without fine-tuning.

desk verdict Solid training recipe for a unified humanoid controller, but the 'complete' claim is under-supported because hidden joints are never scored. read the letter →

arxiv 2502.05641 v1 pith:JK4OV3N7 submitted 2025-02-08 cs.RO cs.AI

Aayam Shrestha , Pan Liu , German Ros , Kai Yuan , Alan Fern This is my paper · ORCID
classification cs.ROcs.AI
keywords humanoidmotiongenerationmultimodaldirectivesimitationlearningphysics-basedcharactercontrolreinforcementmaskeddemonstrationscombinationzero-shotplanning
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 one physics-based controller, trained once, can handle human-motion instructions that arrive incomplete, out of sync, or as mixes of different motions. The proposed Masked Humanoid Controller (MHC) is trained with reinforcement learning on motion-capture demonstrations that are randomly masked, so the policy learns to follow whatever subset of pose information a directive provides and to fill in the rest. Over a dataset of 87 skills, the paper shows the same policy catching up from fallen or mismatched poses, combining upper-body motion from one clip with lower-body motion from another, and completing movements when only a few joints are specified, alongside qualitative demos with VR, video, joystick, and text inputs. The payoff, if true, is a reusable low-level agent that can be directed by any sparse modality and handed to planners for new high-level tasks without any fine-tuning.

What carries the argument

The central object is the masked motion directive, written $d = (\hat{q}_{1:H}, I_{1:H})$, a target motion sequence paired with a binary mask at every time step that says which pose dimensions are constraints. Channel-level masking selects among root height/orientation/velocity, joint rotations, local joint positions, and global keypoint positions; joint-level masking randomly drops a percentage of joints from the keypoint channels. This single representation turns VR, video tracking, joystick, text-to-motion, and full MoCap into the same object, distinguished only by their mask patterns. The training distribution built from these masked directives — augmented by upper/lower-body combinations (dataset $\mathcal{M}^+$), randomly concatenated sub-sequences, random in-plane rotations, and initial poses that include fallen states — is what forces the PPO-trained policy to learn catchup, combine, and complete simultaneously. The reward is a prioritized tracking reward (root height, then orientation, then velocity, then joint positions, each activated only when higher-priority terms are satisfied), a part-wise style reward from an ensemble of five discriminators over body parts, and an energy cost that suppresses jitter.

What would settle it

Run MHC on real Meta Quest 2 headset/controller streams and on video-derived keypoints (e.g., from MeTRAbs) recorded alongside ground-truth motion capture, and compare per-joint error and success rate against the synthetic masked-directive numbers; a large gap would show that the random-mask emulation does not transfer to real input noise.

Watch

Extended reading notes

Core claim

The central claim is that a single multi-objective imitation-learning policy achieves three capabilities no previous motion generator has combined: catch-up, the ability to resynchronize with a target directive from an out-of-sync or fallen state; combine, the ability to imitate directives that blend upper- and lower-body segments from different motion sequences; and complete, the ability to produce the unspecified parts of a motion from sparse directives. The mechanism is a distribution over masked directives: each training episode gives the controller a target motion paired with a binary mask that selects which pose channels and which joints are specified, mimicking whatever a given input modality can supply. The controller is also trained on combinatorial blends of body parts, on randomly concatenated sub-sequences with sharp transitions, and on random initial poses including fallen ones, which together force the three capabilities into one policy. The paper reports that MHC outperforms the ASE baseline on imitation, catchup, and combine metrics on both the training dataset and an ASE-rollout dataset, and demonstrates qualitative multi-modal control plus zero-shot integration with finite-state machines and a data-driven planner for high-level goals.

Load-bearing premise

The random channel-level and joint-level masks used in training are assumed to faithfully represent the incomplete, noisy directives that real VR controllers, video keypoint trackers, joysticks, and text-to-motion systems produce, but the paper only checks the real modalities qualitatively.

Editorial extensions

If this is right

  • A single pretrained controller can be directed by VR, video, joystick, and text inputs without per-modality retraining, since each modality maps to a mask pattern over the same directive representation.
  • New behaviors can be created at runtime by combining upper-body and lower-body directives from different motion clips, without training a separate policy for each combination.
  • Failure recovery and mid-motion switching become a property of the base policy rather than a separate module, since the controller learns to catch up from fallen and out-of-sync states.
  • User-defined high-level tasks can be solved zero-shot by pairing MHC with a data-driven planner or hand-coded FSM, needing only a task reward and a set of valid directives.
  • The training recipe is dataset-agnostic: any MoCap dataset, and any meaningful choice of body-part combination, can be plugged into the masked-directive framework.

Reading between the lines

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

  • If the mask-emulation assumption holds, matching the training mask distribution to actual sensor statistics (occlusion rates, VR tracking dropout, camera viewpoint biases) could further improve real-world transfer, since the paper currently samples masks uniformly at random.
  • The masked-directive abstraction is modality-agnostic in principle, so the same training recipe could extend to other embodiments or to kinematic motion generators, not just physics-based humanoids.
  • The combine capability is only demonstrated for upper/lower-body splits; other compositional splits (left/right limbs, arm/leg pairs, locomotion plus gesture) are natural testable extensions the paper leaves open.
  • Because MHC completes unspecified joints from sparse input, it could serve as a whole-body prior for teleoperation or avatar control from headset-and-controllers alone; the paper shows this qualitatively but does not quantify tracking accuracy under sensor noise.
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

5 major / 6 minor

Summary. The paper introduces the Masked Humanoid Controller (MHC), a physics-based RL policy trained on selectively masked and augmented motion demonstrations. The training uses channel-level and joint-level masking, random poses including falls, concatenated directive subsequences, and a multi-part adversarial style reward. The authors claim that MHC jointly achieves three capabilities: catch-up to out-of-sync directives, combine of upper/lower body segments from different motions, and complete under-specified directives. They evaluate on a dataset of 87 skills against ASE, with quantitative results for imitation, catchup, and combine, and masking experiments for completion, followed by qualitative demos with VR, joystick, video, text, and planning integration. Section 7 states that MHC is the first motion generator to achieve all three capabilities.

Significance. If the claims hold, the contribution is significant: a single physics-based controller that can follow sparse, multi-modal directives and integrate zero-shot with planning would be a useful step for directable virtual characters. The masking-based augmentation is simple and potentially general. However, the quantitative support for the 'complete' capability is currently incomplete, and several comparisons are not apples-to-apples, so the current evidence is weaker than the claims. The qualitative multi-modal demonstrations are encouraging but do not by themselves establish the method's robustness to real sensor noise.

major comments (5)
  1. [Section 6, Metrics paragraph] The paper's own metric definition states that for masked directives, success rate and EMPJPE are measured only on the selected joints. This means the quantitative evidence for the 'complete' capability (Figure 6, Complete paragraph) never evaluates the joints that the model is supposed to fill in. A policy that ignores masked joints while tracking visible joints can score perfectly on these metrics, so the claim that MHC 'completes unspecified parts of motions' is not directly supported. Please report MPJPE and success rate on the masked (unselected) joints, or on the full body, in addition to the visible-joint numbers.
  2. [Section 6, Complete experiment and Figure 6 (right)] The ASE comparison in the joint-masking experiment is not fair: ASE receives fully specified directives and is evaluated on all joints, while MHC receives masked directives and is evaluated only on the unmasked joints. The text even notes that ASE's performance varies across masking levels because 'we only evaluate the metric over unmasked joints.' This makes the apparent superiority of MHC partly an artifact of the evaluation protocol. A meaningful comparison would evaluate both controllers on the same set of joints, e.g., report MHC's full-body MPJPE against ASE's full-body MPJPE, or evaluate both on the same visible subset against the corresponding ground truth.
  3. [Section 6, Combine experiment] The training distribution already includes random upper/lower body combinations from M+ (Section 4.1), and the Combine evaluation samples random pairs of motions from the training and testing sets. It is not stated whether the specific evaluated combinations were excluded from the training augmentation. If they were not, the Combine results largely measure performance on the training distribution rather than generalization to novel combinations. Please specify the hold-out protocol for combinations, and if none was used, add one.
  4. [Table 1 and Figure 6] All quantitative results are reported for a single training run without error bars or multiple seeds. Given the stochasticity of RL training, the observed gaps between MHC and ASE (e.g., Imitation MPJPE 51 vs 123) might not be statistically significant. Please report means and standard deviations over at least three seeds, and indicate the number of evaluation trials used.
  5. [Section 6.2 Multi-Modal Directives] The real-modality demonstrations (VR, joystick, video, text) are qualitative only, using keyframes in Figure 7. No quantitative metrics show how well MHC tracks the actually provided sensor inputs, and no validation is provided that the random mask distribution used in training (uniform random joint/percent masking) matches the noise and sparsity patterns of real VR, MeTRAbs keypoints, or T2M-GPT outputs. Please add quantitative evaluations on these modalities, or at minimum an analysis of the distribution shift between training masks and the masks derived from the real input sources.
minor comments (6)
  1. [Table 1] The word 'expeirments' in the caption should be 'experiments'.
  2. [Section 6.1 Complete] The word 'condider' should be 'consider'.
  3. [Section 4.1 Initial Pose Distribution] The fallen-pose distribution p_fall is mentioned with weight 0.1, but the paper does not describe how these poses are generated or sampled. Please add a brief description.
  4. [Section 6.1 Complete / Figure 6 caption] In Figure 6 (right), the ASE curve varies with masking level, but ASE is never actually given masked directives. The caption and text should clarify that the masking-level axis refers to the evaluation mask, not to ASE's input.
  5. [Section 7] The claim that MHC is 'the first motion generator that achieves all three capabilities' is strong. Since the evidence for 'complete' is currently incomplete, please soften this to 'to our knowledge' and ensure it is supported by the final evaluation.
  6. [Section 5 / Table 2] The heading and go-to-location results (Table 2) are reported without comparison to any baseline or to an upper bound; adding a simple baseline (e.g., random directive selection or a hand-coded FSM) would make the planning integration result more informative.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: CCC capabilities are trained behaviors validated empirically, and the only self-citation (DAC-MDP) is used as a tool, not as load-bearing evidence.

full rationale

The paper's claimed derivation chain is empirical rather than circular. The training distribution (random concatenated subsequences, random initial poses, random channel-level and joint-level masks) is designed to induce catch-up, combine, and complete behaviors, and the paper then evaluates those behaviors on similar distributions. This is a standard train-and-test paradigm, not a reduction of the claims to their inputs by construction. The tracking reward (Eqs. 1-4) uses mask flags m_h, m_o, m_v, m_j to ignore masked channels, but the style reward and energy cost still shape full-body behavior, so the reported Complete metrics are not identical to the training objective. The measurement gap noted in Section 6 — that for under-specified directives the success rate and EMPJPE are "measured only using the selected joints" — weakens the evidence for the Complete capability, but it is an evaluation limitation, not a circular step: no prediction is defined in terms of the fitted quantity, and the style reward provides independent (if unmeasured) pressure on hidden joints. The only self-citation, DAC-MDP [27], is invoked as an external planning framework and is not used to justify the central CCC claims; it is a tool for downstream integration. The qualitative demos with real VR, video, joystick, and text inputs provide external grounding. Therefore no load-bearing circularity is present, and the score is 0.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The scientific content is an empirical training recipe. The hand-set weights, thresholds, mask probabilities, and episode schedule are free parameters in the sense that they are chosen without optimization or ablations that justify them. The method also assumes the simulator is faithful, that random masks approximate real multimodal noise, and that discriminators trained on M generalize. No new physical entities are introduced.

free parameters (6)
  • Tracking/style reward balance = 0.5 / 0.5
    Reward is rt = 0.5 * r_tracking + 0.5 * r_style - energy; weights chosen by hand and no sensitivity analysis is reported.
  • Priority threshold for tracking sub-rewards = 0.9
    Lower-priority terms activate only when higher-priority terms exceed 0.9, per Equations 1-4.
  • Fallen pose mixture weight = 0.1
    Initial pose distribution uses 10% fallen poses to teach catchup behavior.
  • Joint-level masking frequency = 50%
    After channel masking, joint-level masking is applied with probability 0.5 and mask percentage sampled uniformly in [0,100], per Section 4.1.
  • Energy cost coefficients = 0.01 (action diff) and 0.0002 (torque)
    Equation 5; hand-picked to suppress jitter.
  • Episode length and subsequence lengths = L=300; subsequences uniform in 120-240
    Training episodes are 10 seconds with abrupt concatenations, per Section 4.1.
assumptions (5)
  • domain assumption Isaac Gym provides a faithful physics model for the humanoid.
    The claim of physically realistic motions is evaluated only in this simulator; simulator inaccuracies would not be captured.
  • ad hoc to paper Random channel-level and joint-level masks emulate real VR, video, joystick, and text directives.
    Section 4.1 asserts this emulation; no validation against real sensor noise is given, and real-modality results are qualitative.
  • domain assumption The augmented dataset M+, formed by combining upper and lower body subsegments, contains physically plausible directives.
    The Combine capability is trained and tested on such combinations; if arbitrary combinations are unnatural, the learned behavior may not match real user requests.
  • domain assumption Style discriminators trained on the 87-clip dataset provide a valid naturalness reward outside the training distribution.
    The style reward is the average of five discriminators trained only on M; their validity for novel combined or occluded motions is assumed.
  • standard math PPO with fixed entropy and the stated reward weights yields an adequate policy.
    Standard RL practice; hyperparameters are reported only partially, so this remains an assumption for replication.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Generating Physically Realistic and Directable Human Motions from Multi-Modal Inputs." pith.science (2026). https://pith.science/paper/JK4OV3N7

@misc{pith2026250205641,
  author       = {Pith},
  title        = {Pith review of: Generating Physically Realistic and Directable Human Motions from Multi-Modal Inputs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JK4OV3N7}},
  note         = {Machine review of arXiv:2502.05641}
}
read the original abstract

This work focuses on generating realistic, physically-based human behaviors from multi-modal inputs, which may only partially specify the desired motion. For example, the input may come from a VR controller providing arm motion and body velocity, partial key-point animation, computer vision applied to videos, or even higher-level motion goals. This requires a versatile low-level humanoid controller that can handle such sparse, under-specified guidance, seamlessly switch between skills, and recover from failures. Current approaches for learning humanoid controllers from demonstration data capture some of these characteristics, but none achieve them all. To this end, we introduce the Masked Humanoid Controller (MHC), a novel approach that applies multi-objective imitation learning on augmented and selectively masked motion demonstrations. The training methodology results in an MHC that exhibits the key capabilities of catch-up to out-of-sync input commands, combining elements from multiple motion sequences, and completing unspecified parts of motions from sparse multimodal input. We demonstrate these key capabilities for an MHC learned over a dataset of 87 diverse skills and showcase different multi-modal use cases, including integration with planning frameworks to highlight MHC's ability to solve new user-defined tasks without any finetuning.

Figures

Figures reproduced from arXiv: 2502.05641 by the authors.

Figure 1
Figure 1. Showcases generated human motions from multi-modal inputs: (A) VR device, (B) joystick controller, (C) video, and (D) text. Our proposed method, Masked Humanoid Controller (MHC), can generate physically realistic motions from a wide variety of muli-modal directives. demonstrations of real human motion that are relevant to an application. We can then attempt to train a motion generator that can be directed to imitate… view at source ↗
Figure 2
Figure 2. Shows generated motions that illustrate the CCC capabilities. From left to right: MHC is able to generate motions that (1) adjust and catchup starting from an out-of-sync pose, (2) imitate a target directive that combines upper and lower body sub-segments from different motions, and (3) complete the motion from under-specified directives as indicated by missing target outlines. available cues and implicitly filling … view at source ↗
Figure 3
Figure 3. Illustrates the architecture and training details of the MHC framework, which consists of a controller and an ensemble of discriminators. Here the controller is trained to follow an augmented set of masked directives derived from the provided MoCap dataset. The controller gets feedback via tracking objective and style rewards generated by the ensemble of discriminators. Together they enable a directable policy to ge… view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Highlights the potential applications of MHC. [Top] The selective masking of the target directive allows MHC to represent various modalities of motion data under a single framework. These multi-modal inputs include MoCap, full or occluded video, joystick, VR controller…
Figure 5
Figure 5. Figure 5: Illustrates generated motions corresponding to key CCC capabilities of MHC. The simulation (left) displays key-frames of humanoid following different motion directives (right). From top to bottom the simulated humanoid (A) follows an imitation target, (B) transitions f…
Figure 6
Figure 6. Figure 6: (Left) Performance across different channel-level masks. We find that the MHC trained with directive masking retains its imitation performance across different variants of channel masks in contrast to the MHC trained without masking. (Right) Performance across differen…
Figure 7
Figure 7. Figure 7: Illustrates qualitative results using keyframes for motion generation under multi-modal inputs such as (A) VR headset and controllers, (B) joystick controllers, (C) 3D joint positions derived from video and (D) text-to-motion generator. This highlights the versatility …
Figure 8
Figure 8. Figure 8: (A) Key frame visualization of FSMs for Go-To-Location task. FSM A1 (blue) generates a motion of walking towards the goal position while doing right sword slashes and ultimately falling down and recovering on reaching the goal. FSM A2 (green) generates a motion that sw…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

37 extracted references · 26 canonical work pages

  1. [1]

    ACM SIGGRAPH 2023 Conference Proceedings (2023)

    Bae, J., Won, J., Lim, D., Min, C.H., Kim, Y.: Pmp: Learning to physically in- teract with environments using part-wise motion priors. ACM SIGGRAPH 2023 Conference Proceedings (2023)

    work page 2023

  2. [2]

    Micromachines14 (2023)

    Cerón, J.C., Sunny, M.S.H., Brahmi, B., Mendez, L.M., Fareh, R., Ahmed, H.U., Rahman, M.H.: A novel multi-modal teleoperation of a humanoid assistive robot with real-time motion mimic. Micromachines14 (2023)

    work page 2023

  3. [3]

    2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp

    Chen, X., Jiang, B., Liu, W., Huang, Z., Fu, B., Chen, T., Yu, J., Yu, G.: Executing your commands via motion diffusion in latent space. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 18000–18010 (2022)

    work page 2022

  4. [4]

    Proceedings of the 11th ACM SIGGRAPH Conference on Motion, Interaction and Games (2018)

    Chentanez, N., Müller, M., Macklin, M., Makoviychuk, V., Jeschke, S.: Physics- based motion capture imitation with deep reinforcement learning. Proceedings of the 11th ACM SIGGRAPH Conference on Motion, Interaction and Games (2018)

    work page 2018

  5. [5]

    ArXivabs/2309.11351 (2023)

    Dou, Z., Chen, X., Fan, Q., Komura, T., Wang, W.: C·ase: Learning conditional adversarial skill embeddings for physics-based characters. ArXivabs/2309.11351 (2023)

    arXiv 2023

  6. [6]

    2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp

    Du, Y., Kips, R., Pumarola, A., Starke, S., Thabet, A.K., Sanakoyeu, A.: Avatars grow legs: Generating smooth human motion from sparse tracking inputs with diffusion model. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 481–490 (2023)

    work page 2023

  7. [7]

    Ho,J.,Ermon,S.:Generativeadversarialimitationlearning.In:NeuralInformation Processing Systems (2016)

    work page 2016

  8. [8]

    930–939 (2022)

    Huang, Z., Zhou, Z., Chuang, Y.Y., Wu, J., Liu, C.K.: Physically plausible anima- tionofhumanupperbodyfromasingleimage.2023IEEE/CVFWinterConference on Applications of Computer Vision (WACV) pp. 930–939 (2022)

    work page 2022

Show all 37 references
  1. [9]

    In: NeurIPS 2023 Foundation Models for Decision Making Workshop (2023)

    zhengyao jiang, Xu, Y., Wagener, N., Luo, Y., Janner, M., Grefenstette, E., Rock- täschel, T., Tian, Y.: H-GAP: Humanoid control with a generalist planner. In: NeurIPS 2023 Foundation Models for Decision Making Workshop (2023)

    2023

  2. [10]

    SIGGRAPH Asia 2022 Conference Papers (2022)

    Juravsky, J., Guo, Y., Fidler, S., Peng, X.B.: Padl: Language-directed physics- based character control. SIGGRAPH Asia 2022 Conference Papers (2022)

    2022

  3. [11]

    2022 Sixth IEEE International Conference on Robotic Comput- ing (IRC) pp

    Lee, C.H., Yueh, N.C.L., Woo, K.T.: Human-inspired video imitation learning on humanoid model. 2022 Sixth IEEE International Conference on Robotic Comput- ing (IRC) pp. 345–352 (2022)

    2022

  4. [12]

    ACM Transactions on Graphics (TOG)41, 1 – 13 (2022)

    Lee, S., Lee, J., Lee, J.: Learning virtual chimeras by dynamic motion reassembly. ACM Transactions on Graphics (TOG)41, 1 – 13 (2022)

    2022

  5. [13]

    ArXiv abs/2310.04582 (2023)

    Luo, Z.G., Cao, J., Merel, J., Winkler, A., Huang, J., Kitani, K., Xu, W.: Universal humanoid motion representations for physics-based control. ArXiv abs/2310.04582 (2023)

    2023 arXiv

  6. [14]

    Luo, Z., Cao, J., Winkler, A., Kitani, K., Xu, W.: Perpetual humanoid control for real-time simulated avatars.https://doi.org/10.48550/arXiv.2305.06456

    doi arXiv

  7. [15]

    ArXivabs/2206.09286 (2022) 16 A

    Luo, Z., Yuan, Y., Kitani, K.M.: From universal humanoid control to automatic physically valid character creation. ArXivabs/2206.09286 (2022) 16 A. Shrestha et al

    2022 arXiv

  8. [16]

    ArXivabs/2108.10470 (2021)

    Makoviychuk, V., Wawrzyniak, L., Guo, Y., Lu, M., Storey, K., Macklin, M., Hoeller, D., Rudin, N., Allshire, A., Handa, A., State, G.: Isaac gym: High perfor- mance gpu-based physics simulation for robot learning. ArXivabs/2108.10470 (2021)

    2021 arXiv

  9. [17]

    ArXiv abs/1811.11711 (2018)

    Merel, J., Hasenclever, L., Galashov, A., Ahuja, A., Pham, V., Wayne, G., Teh, Y.W., Heess, N.M.O.: Neural probabilistic motor primitives for humanoid control. ArXiv abs/1811.11711 (2018)

    2018 arXiv

  10. [18]

    ArXiv abs/1707.02201 (2017)

    Merel, J., Tassa, Y., Dhruva, T., Srinivasan, S., Lemmon, J., Wang, Z., Wayne, G., Heess, N.M.O.: Learning human behaviors from motion capture by adversarial imitation. ArXiv abs/1707.02201 (2017)

    2017 arXiv

  11. [19]

    Merel, J., Tunyasuvunakool, S., Ahuja, A., Tassa, Y., Hasenclever, L., Pham, V., Erez, T., Wayne, G., Heess, N.: Catch & carry: Reusable neural controllers for vision-guided whole-body tasks

  12. [20]

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

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

    2018

  13. [21]

    Peng, X.B., Guo, Y., Halper, L., Levine, S., Fidler, S.: Ase: Large-scale reusable adversarialskillembeddingsforphysicallysimulatedcharacters.ACMTransactions On Graphics (TOG)41(4), 1–17 (2022)

    2022

  14. [22]

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

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

    2021

  15. [23]

    Reallusion: 3d animation and 2d cartoons made simple

  16. [24]

    Ren, J., Yu, C., Chen, S., Ma, X., Pan, L., Liu, Z.: DiffMimic: Efficient motion mimicking with differentiable physics

  17. [25]

    IEEE Transactions on Biometrics, Behavior, and Identity Science3(1), 16–30 (2021).https://doi.org/ 10.1109/TBIOM.2020.3037257

    Sárándi, I., Linder, T., Arras, K.O., Leibe, B.: MeTRAbs: metric-scale truncation- robust heatmaps for absolute 3D human pose estimation. IEEE Transactions on Biometrics, Behavior, and Identity Science3(1), 16–30 (2021).https://doi.org/ 10.1109/TBIOM.2020.3037257

    2021

  18. [26]

    ArXivabs/1707.06347 (2017)

    Schulman, J., Wolski, F., Dhariwal, P., Radford, A., Klimov, O.: Proximal policy optimization algorithms. ArXivabs/1707.06347 (2017)

    2017 arXiv

  19. [27]

    ArXivabs/2010.08891 (2020)

    Shrestha, A., Lee, S., Tadepalli, P., Fern, A.: Deepaveragers: Offline reinforcement learning by solving derived non-parametric mdps. ArXivabs/2010.08891 (2020)

    2020 arXiv

  20. [28]

    ArXiv abs/2309.11359 (2023)

    Sun, J., Zhang, Q., Duan, Y., Jiang, X., Cheng, C., Xu, R.: Prompt, plan, perform: Llm-based humanoid control via quantized imitation learning. ArXiv abs/2309.11359 (2023)

    2023 arXiv

  21. [29]

    ACM SIG- GRAPH 2023 Conference Proceedings (2023)

    Tessler, C., Kasten, Y., Guo, Y., Mannor, S., Chechik, G., Peng, X.B.: Calm: Conditional adversarial latent models for directable virtual characters. ACM SIG- GRAPH 2023 Conference Proceedings (2023)

    2023

  22. [30]

    org/10.48550/arXiv.2208.07363

    Wagener, N., Kolobov, A., Frujeri, F.V., Loynd, R., Cheng, C.A., Hausknecht, M.: MoCapAct: A multi-task dataset for simulated humanoid control.https://doi. org/10.48550/arXiv.2208.07363

    doi arXiv

  23. [31]

    2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp

    Wang, J., Xu, H., Xu, J., Liu, S., Wang, X.: Synthesizing long-term 3d human motion and interaction in 3d scenes. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) pp. 9396–9406 (2020)

    2020

  24. [32]

    ACM Transactions on Graphics (TOG) 39, 33:1 – 33:12 (2020)

    Won, J., Gopinath, D.E., Hodgins, J.K.: A scalable approach to control diverse be- haviors for physically simulated characters. ACM Transactions on Graphics (TOG) 39, 33:1 – 33:12 (2020)

    2020

  25. [33]

    ACM Transactions on Graphics (TOG)41, 1 – 12 (2022) Generating motions from multi-modal inputs 17

    Won, J., Gopinath, D.E., Hodgins, J.K.: Physics-based character controllers using conditional vaes. ACM Transactions on Graphics (TOG)41, 1 – 12 (2022) Generating motions from multi-modal inputs 17

    2022

  26. [34]

    Proceedings of the ACM on Computer Graphics and Interactive Techniques4, 1 – 22 (2021)

    Xu, P., Karamouzas, I.: A gan-like approach for physics-based imitation learn- ing and interactive control. Proceedings of the ACM on Computer Graphics and Interactive Techniques4, 1 – 22 (2021)

    2021

  27. [35]

    ACM Transactions on Graphics (TOG)42, 1 – 16 (2023)

    Xu, P., Shang, X., Zordan, V.B., Karamouzas, I.: Composite motion learning with task control. ACM Transactions on Graphics (TOG)42, 1 – 16 (2023)

    2023

  28. [36]

    Yao, H., Song, Z., Zhou, Y., Ao, T., Chen, B., Liu, L.: Moconvq: Unified physics- basedmotioncontrolviascalablediscreterepresentations.ArXiv abs/2310.10198 (2023)

    2023 arXiv

  29. [37]

    Head east while facing west

    Zhang, J., Zhang, Y., Cun, X., Zhang, Y., Zhao, H., Lu, H., Shen, X., Shan, Y.: Generating human motion from textual descriptions with discrete representations. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14730–14740 (2023) 18 A. ...

    2023

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.