Half-Physics : Enabling Kinematic 3D Human Model with Physical Interactions
Li Siyao, Yao Feng, Omid Taheri, Chen Change Loy, Michael J. Black · 2025
Read Half-Physics : Enabling Kinematic 3D Human Model with Physical Interactions by Li Siyao, Yao Feng, Omid Taheri, Chen Change Loy, Michael J. Black (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2507-23778.txt.
Introduction
Li Siyao and co-authors add physical touch to a digital human model in a 2025 paper. Popular body models capture shape and pose well but only move points in space. They cannot push objects or rest on surfaces, so virtual people often sink into chairs or pass through props.
Their half-physics method turns posed motion into a live physics simulation while keeping control of the pose. The body can then press, lean, and collide in a believable way. The method needs no training, works for any body shape or motion, and runs in real time. Readers in graphics and robotics get a practical bridge between motion data and physical scenes.
Catalog record
| CREATOR | Li Siyao, Yao Feng, Omid Taheri, Chen Change Loy, Michael J. Black |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 62 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2507-23778
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Half-Physics : Enabling Kinematic 3D Human Model with Physical Interactions Li Siyao, Yao Feng, Omid Taheri, Chen Change Loy, Michael J. Black 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-physic s 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. 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. 1 Introduction Enabling 3D human models to physically interact with their environment is crucial for various applications, such as virtual reality (VR) and embodied AI. Due to the intrinsic nature of existing human parametric models ( i.e. , SMPL [ 20 ] , SMPL-H [ 38 ] and SMPL-X [ 33 ] ) as purely kinematic mesh representations, they lack physical properties, making it difficult for them to interact realistically with the environment. As a result, interaction models often suffer from mesh penetration and physically implausible interactions (see Figure 1 (a)). Although learning-based methods can incorporate loss functions [ 8 , 19 ] or rewards [ 60 ] to suppress these artifacts, they often struggle to generalize. To integrate a 3D human model into a physics simulator, many existing methods employ reinforcement learning (RL) for imitation learning. In this setup, a virtual humanoid agent with joint motors is created inside the simulator. At each simulation timestep \Delta t , an RL policy predicts the torques (forces) for each joint so that the agent’s pose is driven towards the reference kinematic trajectory over time. However, this approach has several drawbacks. (1) Computationally expensive. Imitation learning [ 34 , 35 ] usually requires extensive simulation time to model even a single action. (2) Lack of generalization. Even though recent methods [ 25 , 23 , 59 , 26 ] have improved single-action networks by making them more versatile, they still require separate dedicated training for each motion category ( e.g. , walking). Moreover, for a specific body shape, additional training is necessary to accurately track the reference motion [ 23 ] . (3) Limited fidelity to kinematic data. These methods control motion through per-frame impulses (forces applied over \Delta t ), yet the agent already possesses momentum (velocity). This makes precise control difficult and often results in deviations from the original kinematic data (see Section 4.1 and the supplementary video). 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^{\text{body},i} (for simplicity, we depict only body positions here) at discrete time t , physically intractable within \Delta t , and not able to block penetration. (d) Robotics control uses torque indirectly affect velocity, are physically tractable during the timestep, and able to block penetration. (e) Proposed half physics explicitly infers body velocity towards kinematic target, physically tractable during the timestep, and able to block penetration. In this paper, we propose a simple yet highly efficient and generalizable framework to extend kinematic human models to achieve physically plausible interactions with the environment while maintaining high fidelity to the original kinematic motion. Recall that it is challenging to integrate SMPL-X motions into physical simulations because they are defined by per-frame positions (including both 3D locations and joint rotations), yet these positions remain undefined within the continuous time interval \Delta t . To bridge this gap, rather than using raw position data, we compute the linear velocity of each body part from consecutive frames and explicitly apply it to the body state. This simple yet effective approach enables seamless kinematic-to-physics transfer : if no collision occurs with the environment, the human agent follows the predefined linear velocity to reach the target position in the next frame; if a collision happens, the response is governed by physical laws, ensuring realistic interactions. The proposed approach is advantageous in that it keeps human motion kinematically driven by SMPL-X poses while allowing the surrounding scene and objects to be governed by physics, resulting in physically plausible interactions. We refer such a hybrid control mechanism as “ half physics ”. The proposed half-physics framework requires no training and generalizes effectively to any body shape and motion. Moreover, by directly enforcing velocity instead of torque, it preserves kinematic fidelity while seamlessly integrating physical interactions (see Figure 1 and the video). To evaluate the usability of the proposed half-physics approach, we conduct experiments on human-scene interaction (HSI) and human-object interaction (HOI) separately. For HSI, we verify that half physics eliminates penetration between the generated SMPL-X human body and the static environment. Notably, rather than preventing penetration through motion planning; half physics allows the agent to perform the same motion as in the original kinematic sequence, with penetration naturally handled by the physics simulation (Section 4.2 ). For HOI, we demonstrate that within the half-physics framework, SMPL-X can interact with objects in various ways, including single-handed grasping, two-handed lifting, dropping, and kicking. Additionally, we show that half physics can be used to augment existing interaction datasets by applying different physical factors, enriching interaction outcomes with more diverse physical labels (Section 4.3 ). Across all experiments, half physics takes SMPL-X as input, not only eliminating penetration but also enabling seamless physical interactions, extending the capabilities of kinematic-centric SMPL-X. 2 Related Work Human Parametric Models. The SMPL family [ 20 , 38 , 33 ] provides a series of efficient human body representations to precisely control human shape ( \beta ) and pose ( \theta ) using only a small number of parameters. These models have been widely used in human-centric research, including human pose estimation and motion capture [ 2 , 14 , 15 , 39 , 6 ] , motion generation [ 18 , 43 , 44 , 42 ] , human reconstruction [ 56 , 55 ] , and 3D avatar generation [ 9 ] , enabling the creation of detailed and expressive 3D assets for various content generation tasks. However, SMPL-X remains a purely kinematic model, lacking physical properties necessary for realistic interaction with the environment. Motion Imitation. Motion imitation