Pith. sign in

REVIEW 4 major objections 5 minor 2 cited by

Half-Physics: Enabling Kinematic 3D Human Model with Physical Interactions

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 A kinematic 3D human motion can be made physically interactive with no learning by replacing per-frame poses with per-body-part velocities, so that the body tracks its input exactly until a collision, at which point physics takes over.

desk verdict Useful, simple kinematic-to-physics bridge for SMPL-X, but the PHC+ comparison is circular and the 'physical plausibility' claim outruns the disabled-internal-forces pipeline. read the letter →

arxiv 2507.23778 v2 pith:KUKLD6RC submitted 2025-07-31 cs.CV

Li Siyao , Yao Feng , Omid Taheri , Chen Change Loy , Michael J. Black This is my paper · ORCID
classification cs.CV
keywords halfphysicskinematic-to-physicstransfervelocityoverrideSMPL-Xhuman-sceneinteractionhuman-objectsimulationpenetrationresolution
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

The paper tries to establish that a purely kinematic 3D human motion sequence — positions and joint rotations per frame, with no forces or velocities — can be run inside a physics simulation by converting each frame's target into a velocity. It claims this "half physics" scheme preserves the original motion to machine precision when nothing is touched, lets the simulator resolve collisions and penetration when something is, and needs no training, no reward engineering, and no per-motion or per-body retraining. If true, any kinematic human model used in VR, animation, or embodied AI can gain physically consistent reactions — stopping at walls, lifting, dropping, and kicking objects — purely as an emergent consequence of running the same poses through a physics engine at real-time speed.

What carries the argument

The central mechanism is "half physics": a velocity-override bridge between kinematics and dynamics. Instead of applying joint torques, the method computes the linear velocity of each body part as the finite difference between the current position and the next kinematic target, applies the same finite-difference and spherical-interpolation scheme to global and joint angular velocities, and lets a simulator step the system forward. The body thus moves ballistically toward the kinematic target when nothing touches it, and the engine's collision and friction solver takes over when something does; internal-force and damping terms are switched off because the kinematic input is taken to already carry the muscular coordination, and an optional passive stiffness compensation torque can pull the pose back toward the intended trajectory after contact.

What would settle it

Run a motion-capture experiment in which a human kicks a 0.5 kg ball and a 200 kg ball with the same intended motion, and compare post-contact joint-angle trajectories with half-physics predictions; if the human's joints show muscle-driven braking or pre-impact stiffening that the simulator, which disables internal forces, cannot reproduce, the claim that physical plausibility is preserved after collisions is falsified.

Watch

Extended reading notes

Core claim

The paper's central claim is that a kinematic SMPL-X motion can be embedded in a physics engine with no learning by directly overriding per-body-part velocities instead of driving joints with torques. On dance motions, this achieves tracking error of about 3 micrometers versus centimeters for a trained physical tracker, and 100% success without falls; on scene interactions it reduces penetration from 7.9% of vertices to zero; and on object interactions it produces emergent responses — grasp, lift, drop, kick — that depend correctly on object mass and friction, without any predicted object trajectories. This is presented as a faithful transfer: absent collisions the output equals the kinematic input, and upon collisions the physics engine dictates the interaction.

Load-bearing premise

The load-bearing assumption is that a kinematic motion sequence already encodes the outcome of the body's muscular coordination, so the simulator can safely ignore internal joint forces and damping and respond only to external collisions; if real muscle dynamics matter right after impact, the post-collision body responses are not biomechanically faithful.

Editorial extensions

If this is right

  • Any kinematic SMPL-X motion sequence can be replayed inside a physics engine with zero training, preserving the original poses while blocking penetration against static scenes.
  • Object responses in human-object interaction emerge from mass, friction, and contact rather than from learned or scripted object trajectories, so physical factors can be varied to generate new interaction outcomes.
  • Tracking fidelity shifts from centimeter-scale errors (typical of torque-driven learned trackers) to near machine precision, making the method suitable for applications such as VR where exact alignment with user motion matters.
  • The same half-physics layer can augment interaction datasets with physically realistic labels, such as what happens when a ball is very heavy or a surface is very slippery.
  • Because the conversion is a simple velocity computation, the approach can be ported to other physics simulators that expose per-link velocity interfaces.

Reading between the lines

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

  • If the central claim holds, half physics offers a cheap post-processing layer for generative motion models: sample kinematic poses, then let the simulator decide what happens at contact, with object state as a free byproduct of the interaction.
  • The method's own limitation text notes that internal muscle-skeletal dynamics are not modeled, so its "physical plausibility" is about external-contact behavior, not biomechanical fidelity; a biomechanical validation study would be the decisive test of that scope.
  • The stiffness parameter lambda suggests a controllable trade-off between preserving local joint poses and allowing whole-body displacement; a natural extension would be to fit lambda per motion or per body style from human muscle-tension data.
  • Because the approach does not semantically correct the motion, it would not by itself fix an input that walks into a newly inserted wall; combining it with a motion planner or semantic scene-aware generator is a promising division of labor.
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 introduces "half-physics," a method to embed the kinematic SMPL-X human model into a physics simulator by converting per-frame positions and rotations into body-part velocities. At each timestep, the body parts are assigned the finite-difference velocity toward the next kinematic target, so that in the absence of contact the simulated pose exactly matches the input, while collisions are handled by the physics engine to block penetration and drive object motion. The authors demonstrate the approach on AIST++ dance tracking, penetration removal on the Trumans HSI dataset, and a range of human-object interactions (grasping, lifting, kicking, dropping) with fully simulated object dynamics. The method is learning-free, runs at roughly 950 fps, and generalizes to arbitrary SMPL-X inputs.

Significance. If the claims are taken at face value, the contribution is practically useful: a simple, training-free, real-time bridge between kinematic human motion and physics simulation, with potential applications in VR/AR, embodied AI, and synthetic data augmentation. The paper honestly lists several limitations in Appendix B.3, including the absence of internal actuation dynamics. However, the central claims of "physical plausibility" and "seamless physical interaction" are undermined by two methodological choices: the near-zero tracking error in Table 1 is a tautology of the velocity override, and the explicit disabling of internal force propagation in Appendix A removes the articulated-body response that a real collision would produce. These issues are load-bearing for the paper's main contribution and need to be addressed before the claims can be accepted.

major comments (4)
  1. [Section 4.1, Table 1 and Algorithm 1] The reported MPJPE-g of 0.003 mm is a direct consequence of the control law rather than an empirical tracking result. In Algorithm 1, line 7, the global linear velocity is set to (x_body,t - x_body,t-1)/dt, and joint velocities are set to the angular difference to the target divided by dt. In the absence of collisions, the simulated state must therefore advance exactly to the target pose, up to rounding errors. Comparing this with PHC+, whose agent must contend with inertia and internal dynamics, is an apples-to-oranges comparison. The claim that half-physics reduces tracking error from centimeter scale to negligible is misleading; the correct statement is that the output is identical to the input when no contact occurs. Please reframe this comparison or report fidelity only under contacts.
  2. [Appendix A, 'Damping-Free Dynamics and the Neglect of Internal Force'; Table 4] Disabling internal force propagation in the Featherstone pipeline eliminates the inertial coupling between links. In a real articulated collision, the contact impulse at the foot is resisted by the effective mass of the whole leg (shank, thigh, pelvis), and joint reaction forces transmit deceleration through the skeleton. With internal forces disabled, the post-impact state of the struck link depends almost entirely on the link's own inertia, and the next-frame velocity override then resets the joint velocities toward the kinematic target. As a result, examples such as Figure 5(h) (200 kg ball) and Figure 5(i) (200 kg suitcase) demonstrate collision-local deformation rather than whole-body physical response, and the object's exit speed may be incorrect even when the visual outcome looks plausible. The paper's own Table 4 shows that enabling internal forces destroys the perfect tracking, indicating that the fidelity claim is purchased exactly by suppressing the dynamics needed for a faithful articulated collision response. I recommend either (i) a quantitative validation of contact impulses and object velocities against a full rigid-body simulator with internal forces enabled, or (ii) a careful restatement of the contribution as 'kinematically driven collision response with physically simulated object dynamics,' with the human body's physical plausibility explicitly qualified.
  3. [Appendix A, Eq. (4) and Section 4.3] Passive Joint Stiffness Compensation (PJSC) applies a corrective joint torque, which is itself an internal actuation. The main HOI experiments in Section 4.3 use lambda = 0 (no PJSC), so the claimed 'fingers bend under strain' in Figure 5(i) cannot be attributed to passive stiffness; it must be a result of the velocity override and contact dynamics within the timestep, whose physical meaning is unclear. The paper should clarify what physical mechanism produces the bending, and how PJSC (which re-introduces internal torque) interacts with the 'no internal forces' setting and with the reported penetration and tracking metrics.
  4. [Section 4.2 and Figure 4] The penetration elimination results are plausible and well presented, but the evaluation only reports zero penetration and does not quantify side effects such as foot sliding, body rotation, or deviation from the intended motion after contacts. Appendix B.3 mentions foot sliding as a possible issue, but the main body of the paper claims the motion remains faithful to the original intent without qualification. A quantitative measure of motion deviation after contact (e.g., global trajectory displacement or joint-angle deviation) would strengthen the HSI evaluation and clarify the trade-off between penetration removal and kinematic fidelity.
minor comments (5)
  1. [Section 4.1] The text says 'The results are shown in Table 2' but the comparison results appear in Table 1; Table 2 has the HSI results. Please correct the references.
  2. [Figure 2 caption] The caption contains the typo 'frameword' instead of 'framework'.
  3. [Section 3.2] The notation introduces both global body velocity and per-part velocities, but the superscript convention is used inconsistently; for instance, ˙x_body,0_t is used for the root global velocity, while later in the text the same symbol with index i denotes part velocities. Please clarify the indexing in the equations.
  4. [Figure 8] PJSC is defined only in Appendix A but is referenced in the main text and Figure 8. Consider defining the acronym at first use or moving a short explanation to Section 3.
  5. [References] Reference [38] cites the SMPL-H/Embodied Hands model as an arXiv preprint; the published version (ACM Transactions on Graphics, 2022) should be cited if available.

Circularity Check

1 steps flagged · score 6.0 of 10

The claimed near-perfect kinematic tracking (MPJPE 3 µm) is definitional: velocities are set to the finite difference to the target pose, so the no-collision output equals the input by construction; the contact-response contribution is genuinely emergent and not circular.

  1. self definitional [Section 3.2, Algorithm 1 (lines 6-8); Section 4.1, Table 1]
    "Joint velocities ˙qjoint t ... are computed via spherical linear interpolation ... between their targets (qjoint t and qbody,0 t ) and their current states (ˆqjoint t−1 and ˆqbody,0 t−1 ), respectively. In contrast, the global linear velocity ˙xbody,0 t is computed as the difference between the target global positions at frames t and t − 1. ... If no collision occurs, the half-physics SMPL-X variant can maintain the same pose sequence as the original SMPL-X, i.e., ˆθ = H(θ). ... HP achieves a global error of merely 3 µm ..."

    The velocity command is, by the algorithm's definition, exactly the displacement from the current simulated state to the kinematic target divided by Δt, with rotations handled by SLERP toward the same target. Integrating that prescribed velocity over one timestep reproduces the target pose whenever no collision occurs. Therefore the 'tracking' output is the kinematic input by construction, and the reported MPJPE of 3 µm is an identity check on the finite-difference formula rather than an empirical tracking result. The paper itself states that in the no-collision case the output equals the input and that the residual is rounding error. Comparing this number with PHC+'s centimeter-level error conflates a definitional identity with a control achievement.

full rationale

The paper's central practical novelty is the half-physics velocity override, and the object-response experiments (kicking, lifting, dropping, mass/friction variation) are emergent from the simulator rather than read off from the input. However, the headline quantitative claim of 3 µm tracking fidelity is tautological: the per-frame linear and angular velocities are defined as the finite differences (or SLERP differences) between the current state and the kinematic target, so integrating them with no contact returns the target. The paper explicitly says the no-collision output is identical to the input and that the residual is rounding. This makes the MPJPE comparison to PHC+ a comparison between a definitional identity and a genuine physical-tracking result, which is a real circular step in the evaluation. No load-bearing self-citation chain or imported uniqueness theorem is present; the Appendix A decision to disable internal force propagation is a stated modeling assumption and limitation rather than a circular step, and Appendix B.3 candidly admits that internal actuating dynamics are not modeled. Because one central supporting result reduces by construction while the main interaction contribution remains independent, the appropriate score is 6.

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

The approach is simple and uses standard physics simulation, but it relies on a handful of manually chosen gains (lambda, N) and on the strong assumption that internal body dynamics can be ignored. Body mass, inertia, and friction values are not fully specified, adding hidden degrees of freedom.

free parameters (4)
  • PJSC gain lambda = 0 (most experiments), 1 (HSI)
    Controls strength of corrective torque that resists collision-induced pose changes; tuned per experiment.
  • PJSC substep count N = 8 (HSI)
    Number of substeps used to apply PJSC torque; set to 8 in HSI, not specified elsewhere.
  • Body part masses and inertias = not specified
    Human body mass and inertia of the 55-rigid-part SMPL-X variant are not reported; presumably default or hand-set values, which affect contact response.
  • Contact friction coefficients = not specified
    Friction between the human body and objects or environment is not reported except in the object-friction ablation where object friction is 0.1; the human-side friction is implicit.
assumptions (4)
  • domain assumption SMPL-X body can be approximated as 55 rigid body parts derived from vertex segmentation
    Used to build the collision and simulation model; if the rigid parts do not match the SMPL-X surface, contact points and penetration metrics could be inaccurate. See Section 3.2.
  • domain assumption Internal joint damping and internal force propagation can be disabled because the kinematic motion already encodes complete muscle-skeletal coordination
    This is the load-bearing modeling assumption; without it the system would deviate from the kinematic reference, but with it the post-collision response lacks realistic inertial coupling. Appendix A.
  • standard math Finite-difference velocity over Delta t accurately represents the intended body motion during the simulation step
    Assumes that linear interpolation of positions (and SLERP for rotations) is a valid approximation of the body's motion within the frame interval, which is standard and acceptable for high frame rates.
  • domain assumption The physics engine's collision detection with the rigid variant is numerically consistent with mesh-level penetration measurements
    The paper reports zero penetration after simulation, but the metric is computed on SMPL-X meshes, not on the variant; small discrepancies between variant and mesh could hide shallow contact.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Half-Physics: Enabling Kinematic 3D Human Model with Physical Interactions." pith.science (2026). https://pith.science/paper/KUKLD6RC

@misc{pith2026250723778,
  author       = {Pith},
  title        = {Pith review of: Half-Physics: Enabling Kinematic 3D Human Model with Physical Interactions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KUKLD6RC}},
  note         = {Machine review of arXiv:2507.23778}
}
read the original abstract

While current general-purpose 3D human models (e.g., SMPL-X) efficiently represent accurate human shape and pose, they lacks the ability to physically interact with the environment due to the kinematic nature. As a result, kinematic-based interaction models often suffer from issues such as interpenetration and unrealistic object dynamics. To address this limitation, we introduce a novel approach that embeds SMPL-X into a tangible entity capable of dynamic physical interactions with its surroundings. Specifically, we propose a "half-physics" mechanism that transforms 3D kinematic motion into a physics simulation. Our approach maintains kinematic control over inherent SMPL-X poses while ensuring physically plausible interactions with scenes and objects, effectively eliminating penetration and unrealistic object dynamics. Unlike reinforcement learning-based methods, which demand extensive and complex training, our half-physics method is learning-free and generalizes to any body shape and motion; meanwhile, it operates in real time. Moreover, it preserves the fidelity of the original kinematic motion while seamlessly integrating physical interactions

Figures

Figures reproduced from arXiv: 2507.23778 by the authors.

Figure 1
Figure 1. Kinematic and half-physics interactions of the SMPL-X 3D human model with the environment. (a) Standard kinematic SMPL-X meshes (pink) do not account for physical interactions, resulting in implausible penetration through obstacles. (b) Our proposed half-physics framework (blue) allows the human model to produce physical interactions while keeping the original kinematic control. Abstract While current general-purpos… view at source ↗
Figure 2
Figure 2. Illustration of the proposed half physics frameword. Half physics bridges kinematics and physics by assigning equivalent velocities, enabling physically plausible transitions from purely kinematic inputs. (a) Shape-aware articulated variant (blue) of original human model (pink). (b) Two frames with kinematic poses, with penetration at fr. t+1. (c) Kinematic poses x body,i (for simplicity, we depict only body positio… view at source ↗
Figure 3
Figure 3. Qualitative comparison to PHC+ [24]. The tracked subject falls over PHC+ (red box), while ours keeps high fidelity to kinematic reference [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Qualitative results of Trumans [13] (pink) and Trumans+HP (ours) (blue). The proposed half physics can effectively eliminate different kinds of penetration in human scene interactions while maintain high fidelity to original kinematic motion. local MPJPE, which conside…
Figure 5
Figure 5. Figure 5: Half physics can enable kinematic human models to perform diverse interactions with objects. (a-e) Human motions are kinematic while objects (orange) are driven by physics. (f) Without contact, pure kinematic models generate unrealistic object (gray) moving, while half…
Figure 6
Figure 6. Figure 6: Comparison on original AI Habitat [36] and our implemented half physics (HP) [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 8
Figure 8. Figure 8: Passive Joint Stiffness Compensation (PJSC) in half physics. (a, b) Humans can actively regulate how easily their pose changes in response to external forces by relaxing or tensing (red) their muscles, even under the same kinematic pose. (c) In contrast, simulated arti…
Figure 9
Figure 9. Figure 9: Qualitative results of Trumans [13] (pink) and Trumans+HP (ours) (cyan). The proposed half physics can effectively eliminate different kinds of penetration in human scene interactions. Social Impact. All of the aforementioned examples demonstrate the potential for posi…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. HumanCLAW: Can Vision-Language Models Act Through a Body?

    cs.CV 2026-07 conditional novelty 7.0 of 10

    Off-the-shelf VLMs fail closed-loop whole-body find-navigate-interact tasks (best 16.8%) because they lack embodied self-awareness, not target recognition.

  2. Advances in 4D Representation: Geometry, Motion, and Interaction

    cs.CV 2025-10 conditional novelty 4.0 of 10

    A representation-centric survey of 4D generation and reconstruction, organized by geometry, motion, and interaction, with qualitative trade-off comparisons across seven representation families.

Reference graph

Works this paper leans on

59 extracted references · 53 canonical work pages · cited by 2 Pith papers

  1. [1]

    Karen Liu

    Joao Pedro Araujo, Jiaman Li, Karthik Vetrivel, Rishi Agarwal, Deepak Gopinath, Jiajun Wu, Alexander Clegg, and C. Karen Liu. CIRCLE: Capture in rich contextual environments. In CVPR, 2023

    work page 2023

  2. [2]

    Keep it smpl: Automatic estimation of 3d human pose and shape from a single image

    Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J Black. Keep it smpl: Automatic estimation of 3d human pose and shape from a single image. In ECCV, 2016

    work page 2016

  3. [3]

    Bullet physics simulation

    Erwin Coumans. Bullet physics simulation. In ACM SIGGRAPH 2015 Courses, page 1. ACM, 2015

    work page 2015

  4. [4]

    Habitat 2.0: Training home assistants to rearrange their habitat

    Mikael Dallaire Côté, Eric Undersander, Alexander Clegg, et al. Habitat 2.0: Training home assistants to rearrange their habitat. In NeurIPS, pages 21525–21537, 2021

    work page 2021

  5. [5]

    The calculation of robot dynamics using articulated-body inertias

    Roy Featherstone. The calculation of robot dynamics using articulated-body inertias. Interna- tional Journal of Robotics Research, 2(1):13–30, 1983

    work page 1983

  6. [6]

    Chatpose: Chatting about 3d human pose

    Yao Feng, Jing Lin, Sai Kumar Dwivedi, Yu Sun, Priyanka Patel, and Michael J Black. Chatpose: Chatting about 3d human pose. In CVPR, 2024

    work page 2024

  7. [7]

    Mohamed Hassan, Duygu Ceylan, Ruben Villegas, Jun Saito, Jimei Yang, Yi Zhou, and Michael J. Black. Stochastic scene-aware motion prediction. In ICCV, 2021

    work page 2021

  8. [8]

    Resolving 3d human pose ambiguities with 3d scene constraints

    Mohamed Hassan, Vasileios Choutas, Dimitrios Tzionas, and Michael J Black. Resolving 3d human pose ambiguities with 3d scene constraints. In ICCV, 2019

    work page 2019

Show all 59 references
  1. [9]

    Avatarclip: Zero-shot text-driven generation and animation of 3d avatars

    Fangzhou Hong, Mingyuan Zhang, Liang Pan, Zhongang Cai, Lei Yang, and Ziwei Liu. Avatarclip: Zero-shot text-driven generation and animation of 3d avatars. arXiv preprint arXiv:2205.08535, 2022

    2022 arXiv

  2. [10]

    Diffusion-based generation, optimization, and planning in 3d scenes

    Siyuan Huang, Zan Wang, Puhao Li, Baoxiong Jia, Tengyu Liu, Yixin Zhu, Wei Liang, and Song-Chun Zhu. Diffusion-based generation, optimization, and planning in 3d scenes. In ICCV, 2023

    2023

  3. [11]

    Advanced character physics

    Thomas Jakobsen. Advanced character physics. In Game developers conf., 2001. 10

    2001

  4. [12]

    Au- tonomous character-scene interaction synthesis from text instruction

    Nan Jiang, Zimo He, Zi Wang, Hongjie Li, Yixin Chen, Siyuan Huang, and Yixin Zhu. Au- tonomous character-scene interaction synthesis from text instruction. In SIGGRAPH Asia, 2024

    2024

  5. [13]

    Scaling up dynamic human-scene interaction modeling

    Nan Jiang, Zhiyuan Zhang, Hongjie Li, Xiaoxuan Ma, Zan Wang, Yixin Chen, Tengyu Liu, Yixin Zhu, and Siyuan Huang. Scaling up dynamic human-scene interaction modeling. In CVPR, 2024

    2024

  6. [14]

    End-to-end recovery of human shape and pose

    Angjoo Kanazawa, Michael J Black, David W Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In CVPR, 2018

    2018

  7. [15]

    Vibe: Video inference for human body pose and shape estimation

    Muhammed Kocabas, Nikos Athanasiou, and Michael J Black. Vibe: Video inference for human body pose and shape estimation. In CVPR, 2020

    2020

  8. [16]

    Karen Liu

    Jiaman Li, Alexander Clegg, Roozbeh Mottaghi, Jiajun Wu, Xavier Puig, and C. Karen Liu. Controllable human-object interaction synthesis. In ECCV, 2024. Oral Presentation

    2024

  9. [17]

    Karen Liu

    Jiaman Li, Jiajun Wu, and C. Karen Liu. Object motion guided human motion synthesis. ACM ToG, 2023

    2023

  10. [18]

    Ai choreographer: Music conditioned 3d dance generation with aist++

    Ruilong Li, Shan Yang, David A Ross, and Angjoo Kanazawa. Ai choreographer: Music conditioned 3d dance generation with aist++. In ICCV, 2021

    2021

  11. [19]

    Revisit human-scene interaction via space occupancy

    Xinpeng Liu, Haowen Hou, Yanchao Yang, Yong-Lu Li, and Cewu Lu. Revisit human-scene interaction via space occupancy. In ECCV, 2024

    2024

  12. [20]

    SMPL: A skinned multi-person linear model

    Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. SMPL: A skinned multi-person linear model. ACM ToG, 34(6), 2015

    2015

  13. [21]

    Carl: Controllable agent with reinforcement learning for quadruped locomotion

    Ying-Sheng Luo, Jonathan Hans Soeseno, Trista Pei-Chun Chen, and Wei-Chao Chen. Carl: Controllable agent with reinforcement learning for quadruped locomotion. ACM ToG, 39(4):38– 1, 2020

    2020

  14. [22]

    Omnigrasp: Grasping diverse objects with simulated humanoids

    Zhengyi Luo, Jinkun Cao, Sammy Christen, Alexander Winkler, Kris Kitani, and Weipeng Xu. Omnigrasp: Grasping diverse objects with simulated humanoids. In NeurIPS, 2024

    2024

  15. [23]

    Perpetual humanoid control for real-time simulated avatars

    Zhengyi Luo, Jinkun Cao, Kris Kitani, Weipeng Xu, et al. Perpetual humanoid control for real-time simulated avatars. In ICCV, 2023

    2023

  16. [24]

    Universal humanoid motion representations for physics-based control

    Zhengyi Luo, Jinkun Cao, Josh Merel, Alexander Winkler, Jing Huang, Kris M Kitani, and Weipeng Xu. Universal humanoid motion representations for physics-based control. In ICLR, 2024

    2024

  17. [25]

    Embodied scene-aware human pose estimation

    Zhengyi Luo, Shun Iwase, Ye Yuan, and Kris Kitani. Embodied scene-aware human pose estimation. NeurIPS, 2022

    2022

  18. [27]

    Humanoidlympics: Sports environments for physically simulated humanoids

    Zhengyi Luo, Jiashun Wang, Kangni Liu, Haotian Zhang, Chen Tessler, Jingbo Wang, Ye Yuan, Jinkun Cao, Zihui Lin, Fengyi Wang, Jessica Hodgins, and Kris Kitani. Humanoidlympics: Sports environments for physically simulated humanoids. In arXiv preprint arXiv:2407.00187, 2025

    2025 arXiv

  19. [28]

    Xpbd: position-based simulation of compliant constrained dynamics

    Miles Macklin, Matthias Müller, and Nuttapong Chentanez. Xpbd: position-based simulation of compliant constrained dynamics. In ICMG, 2016

    2016

  20. [29]

    Amass: Archive of motion capture as surface shapes

    Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. Amass: Archive of motion capture as surface shapes. In ICCV, 2019

    2019

  21. [30]

    Generating continual human motion in diverse 3d scenes

    Aymen Mir, Xavier Puig, Angjoo Kanazawa, and Gerard Pons-Moll. Generating continual human motion in diverse 3d scenes. In ICCV, 2023. 11

    2023

  22. [31]

    Detailed rigid body simulation with extended position based dynamics

    Matthias Müller, Miles Macklin, Nuttapong Chentanez, Stefan Jeschke, and Tae-Yong Kim. Detailed rigid body simulation with extended position based dynamics. In CGF, 2020

    2020

  23. [32]

    Synthesizing physically plausible human motions in 3d scenes

    Liang Pan, Jingbo Wang, Buzhen Huang, Junyu Zhang, Haofan Wang, Xu Tang, and Yangang Wang. Synthesizing physically plausible human motions in 3d scenes. In 3DV, 2021

    2021

  24. [33]

    Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In CVPR, 2019

    2019

  25. [34]

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

    Xue Bin Peng, Pieter Abbeel, Sergey Levine, and Michiel Van de Panne. Deepmimic: Example- guided deep reinforcement learning of physics-based character skills. ACM ToG, 37(4):1–14, 2018

    2018

  26. [35]

    Amp: Adversarial motion priors for stylized physics-based character control

    Xue Bin Peng, Ze Ma, Pieter Abbeel, Sergey Levine, and Angjoo Kanazawa. Amp: Adversarial motion priors for stylized physics-based character control. ACM ToG, 40(4):1–20, 2021

    2021

  27. [36]

    Habitat 3.0: A co-habitat for humans, avatars, and robots

    Xavier Puig, Eric Undersander, Andrew Szot, et al. Habitat 3.0: A co-habitat for humans, avatars, and robots. arXiv preprint arXiv:2310.13724, 2023

    2023 arXiv

  28. [37]

    Diffmimic: Efficient motion mimicking with differentiable physics

    Jiawei Ren, Cunjun Yu, Siwei Chen, Xiao Ma, Liang Pan, and Ziwei Liu. Diffmimic: Efficient motion mimicking with differentiable physics. In ICLR, 2022

    2022

  29. [38]

    Embodied hands: Modeling and capturing hands and bodies together

    Javier Romero, Dimitrios Tzionas, and Michael J Black. Embodied hands: Modeling and capturing hands and bodies together. arXiv preprint arXiv:2201.02610, 2022

    2022 arXiv

  30. [39]

    Airpose: Multi-view fusion network for aerial 3d human pose and shape estimation

    Nitin Saini, Elia Bonetto, Eric Price, Aamir Ahmad, and Michael J Black. Airpose: Multi-view fusion network for aerial 3d human pose and shape estimation. IEEE Robotics and Automation Letters, 7(2):4805–4812, 2022

    2022

  31. [40]

    Habitat: A platform for embodied ai research

    Manolis Savva, Abhishek Kadian, Oleksandr Maksymets, et al. Habitat: A platform for embodied ai research. In ICCV, 2019

    2019

  32. [41]

    Animating rotation with quaternion curves

    Ken Shoemake. Animating rotation with quaternion curves. In Annual Conf. on Comput. Graph. and Interactive Techs, 1985

    1985

  33. [42]

    Duolando: Follower gpt with off-policy reinforcement learning for dance accompaniment

    Li Siyao, Tianpei Gu, Zhitao Yang, Zhengyu Lin, Ziwei Liu, Henghui Ding, Lei Yang, and Chen Change Loy. Duolando: Follower gpt with off-policy reinforcement learning for dance accompaniment. In ICLR, 2024

    2024

  34. [43]

    Bailando: 3d dance generation by actor-critic gpt with choreographic memory

    Li Siyao, Weijiang Yu, Tianpei Gu, Chunze Lin, Quan Wang, Chen Qian, Chen Change Loy, and Ziwei Liu. Bailando: 3d dance generation by actor-critic gpt with choreographic memory. In CVPR, 2022

    2022

  35. [44]

    Bailando++: 3d dance gpt with choreographic memory

    Li Siyao, Weijiang Yu, Tianpei Gu, Chunze Lin, Quan Wang, Chen Qian, Chen Change Loy, and Ziwei Liu. Bailando++: 3d dance gpt with choreographic memory. IEEE TPAMI, 45(12):14192–14207, 2023

    2023

  36. [45]

    Black, and Dimitrios Tzionas

    Omid Taheri, Nima Ghorbani, Michael J. Black, and Dimitrios Tzionas. GRAB: A dataset of whole-body human grasping of objects. In ECCV, 2020

    2020

  37. [46]

    Bermano, and Michiel van de Panne

    Guy Tevet, Sigal Raab, Setareh Cohan, Daniele Reda, Zhengyi Luo, Xue Bin Peng, Amit H. Bermano, and Michiel van de Panne. CLoSD: Closing the loop between simulation and diffusion for multi-task character control. In ICLR, 2025

    2025

  38. [47]

    Mocapact: A multi-task dataset for simulated humanoid control.NeurIPS, 2022

    Nolan Wagener, Andrey Kolobov, Felipe Vieira Frujeri, Ricky Loynd, Ching-An Cheng, and Matthew Hausknecht. Mocapact: A multi-task dataset for simulated humanoid control.NeurIPS, 2022

    2022

  39. [48]

    Synthesizing long-term 3d human motion and interaction in 3d scenes

    Jiashun Wang, Huazhe Xu, Jingwei Xu, Sifei Liu, and Xiaolong Wang. Synthesizing long-term 3d human motion and interaction in 3d scenes. In CVPR, 2021

    2021

  40. [49]

    Towards diverse and natural scene-aware 3d human motion synthesis

    Jingbo Wang, Yu Rong, Jingyuan Liu, Sijie Yan, Dahua Lin, and Bo Dai. Towards diverse and natural scene-aware 3d human motion synthesis. In CVPR, 2022. 12

    2022

  41. [50]

    Scene-aware generative network for human motion synthesis

    Jingbo Wang, Sijie Yan, Bo Dai, and Dahua Lin. Scene-aware generative network for human motion synthesis. In CVPR, 2021

    2021

  42. [51]

    PhysHOI: Physics-based imitation of dynamic human-object interaction

    Yinhuai Wang, Jing Lin, Ailing Zeng, Zhengyi Luo, Jian Zhang, and Lei Zhang. PhysHOI: Physics-based imitation of dynamic human-object interaction. arXiv preprint arXiv:2312.04393, 2023

    2023 arXiv

  43. [52]

    SkillMimic: Learning reusable basketball skills from demonstrations

    Yinhuai Wang, Qihan Zhao, Runyi Yu, Ailing Zeng, Jing Lin, Zhengyi Luo, Hok Wai Tsui, Jiwen Yu, Xiu Li, Qifeng Chen, Jian Zhang, Lei Zhang, and Ping Tan. SkillMimic: Learning reusable basketball skills from demonstrations. In CVPR, 2025

    2025

  44. [53]

    Karen Liu

    Zhen Wu, Jiaman Li, Pei Xu, and C. Karen Liu. Human-object interaction from human-level instructions. arXiv preprint, arXiv 2406.17840, 2024

    2024 arXiv

  45. [54]

    Unified human-scene interaction via prompted chain-of-contacts

    Zeqi Xiao, Tai Wang, Jingbo Wang, Jinkun Cao, Wenwei Zhang, Bo Dai, Dahua Lin, and Jiangmiao Pang. Unified human-scene interaction via prompted chain-of-contacts. In ICLR, 2024

    2024

  46. [55]

    Econ: Explicit clothed humans optimized via normal integration

    Yuliang Xiu, Jinlong Yang, Xu Cao, Dimitrios Tzionas, and Michael J Black. Econ: Explicit clothed humans optimized via normal integration. In CVPR, 2023

    2023

  47. [56]

    Icon: Implicit clothed humans obtained from normals

    Yuliang Xiu, Jinlong Yang, Dimitrios Tzionas, and Michael J Black. Icon: Implicit clothed humans obtained from normals. In CVPR, 2022

    2022

  48. [57]

    Interdiff: Generating 3d human-object interactions with physics-informed diffusion

    Sirui Xu, Zhengyuan Li, Yu-Xiong Wang, and Liang-Yan Gui. Interdiff: Generating 3d human-object interactions with physics-informed diffusion. In ICCV, 2023

    2023

  49. [58]

    Intermimic: Towards universal whole-body control for physics-based human-object interactions

    Sirui Xu, Hung Yu Ling, Yu-Xiong Wang, and Liangyan Gui. Intermimic: Towards universal whole-body control for physics-based human-object interactions. In CVPR, 2025

    2025

  50. [59]

    Learning physically simulated tennis skills from broadcast videos

    Ye Yuan, Viktor Makoviychuk, Y Guo, S Fidler, XB Peng, and K Fatahalian. Learning physically simulated tennis skills from broadcast videos. ACM ToG, 42(4), 2023

    2023

  51. [60]

    internal

    Kaifeng Zhao, Yan Zhang, Shaofei Wang, Thabo Beeler, and Siyu Tang. DIMOS: Synthesizing diverse human motions in 3d indoor scenes. In ICCV, 2023. A Implementation Details of Half-Physics Simulator In theory, half physics can be extended to any simulator. However, since half ph...

    2023

Pith tools

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