We introduce an efficient solution to the problem of continuous collision detection (CCD) between triangle geometry and signed distance fields (SDFs). We formulate the triangle-SDF collision problem as a novel spatio-temporal local optimization that solves for the first time of impact between a triangle and an SDF isosurface. Our method offers improved robustness over point sampling methods, and outperforms recent triangle-SDF discrete collision detection (DCD) algorithms. Furthermore, a novel method for adaptively refining the potential collision points on large triangles is proposed for robust triangle-SDF collision detection with coarse meshes. This enables the use of reduced geometry for efficient simulations. We demonstrate the benefits of our approach by comparing to state-of-the-art algorithms for triangle-SDF collision detection, and showcase its effectiveness through simulations involving complex collision scenarios.

Physics-based characters need to plan their movements over-and-around obstacles. Diffusion models offer a possible solution, as they allow a motion planner to sample from a potentially diverse distribution of possible futures. However, they may also generate flawed plans because some samples may not satisfy hard constraints that exist only implicitly in the training data, e.g., guaranteeing balance or obstacle clearance. We propose learned viability filters that can efficiently predict the future success of a given plan, i.e., diffusion sample, and thereby enforce an implicit future-success constraint. Multiple viability filters can also be composed together at run-time to take multiple potential constraints into consideration. We demonstrate the approach on detailed footstep planning for 3D human locomotion tasks, showing the effectiveness of the viability filters in performing online planning for box-climbing, step-over walls, and obstacle avoidance. We compare to a number of alternative planning methods including reinforcement learning and return-conditioned diffusion, and further show that using viability filters is significantly faster than guidance-based diffusion prediction.

Ants exhibit unique abilities to self-assemble into animate, living structures. Such structures display properties of both fluid and solid-like, deformable materials. Despite much progress in our understanding of ant aggregation dynamics, simulating such phenomena has been largely overlooked in real-time graphics and animation applications. We present a constraints-based approach for simulating the collective dynamics of ants interactively with compelling physical realism. Through several experiments and comparisons, we demonstrate that our method can efficiently capture the underlying physical rules of such aggregations, which can inform further research into living structures that can morph and self-repair.

We propose a unified volumetric combustion simulator that supports general wooden structures capturing the multi-phase combustion of charring materials. Complex geometric structures are represented in a voxel grid for the efficient evaluation of volumetric effects. In addition, a signed distance field is introduced to query the surface information required to compute the insulating effect caused by the char layer. Non-charring materials such as acrylic glass or non-combustible materials such as stone can also be modeled in the simulator. Adaptive data structures are utilized to enable memory-efficient computations within our multiresolution approach. The simulator is qualitatively validated by showcasing the numerical simulation of a variety of scenes covering different kinds of structural configurations and materials. Two-way coupling of our combustion simulator and position-based dynamics is demonstrated capturing characteristic mechanical deformations caused by the combustion process. The volumetric combustion process of wooden structures is further quantitatively assessed by comparing our simulated results to sub-surface measurements of a real-world combustion experiment.

Performing everyday tasks requires both large-scale body movements driven by the arms, legs, and torso, and fine motor skills, particularly in the hands. However, existing reinforcement learning approaches often struggle to efficiently acquire these diverse motion skills from scratch, leading to slow convergence or suboptimal policies. We observed that the motor skills of distinct body parts exhibit a certain level of independence. This suggests the potential advantage of independently pretraining specific dexterous limbs prior to their integration in full-body motion tasks. Inspired by this, we present Part-wise Heterogeneous Agents (PHA) , a cooperative multi-agent reinforcement learning approach where body parts are treated as independent agents, allowing for specialized skill acquisition and cooperative execution of complex full-body tasks. Furthermore, our method enables the pretraining of fine motor skills, such as gripping a bar or grabbing a climbing hold, before integrating them with other body parts for complex whole-body coordination, thus introducing part-wise Reusable Policy Priors . We tested our technique on challenging tasks such as rope climbing, rock bouldering and traversing a horizontal ladder. Our approach not only accelerates convergence, but also improves overall policy quality, achieving motion tasks that single-agent approaches struggle to solve. Our results also demonstrate adaptability, enabling Reusable Policy Priors to adjust their policies to successfully perform complex tasks in scenarios not seen during training.

Modeling the dynamic behavior of deformable objects is crucial for creating realistic digital worlds. While conventional simulations produce high-quality motions, their computational costs are often prohibitive. Subspace simulation techniques address this challenge by restricting deformations to a lower-dimensional space, improving performance while maintaining visually compelling results. However, even subspace methods struggle to meet the stringent performance demands of portable devices such as virtual reality headsets and mobile platforms. To overcome this limitation, we introduce a novel subspace simulation framework powered by a neural latent-space integrator. Our approach leverages self-supervised learning to enhance inference stability and generalization. By operating entirely within latent space, our method eliminates the need for full-space computations, resulting in a highly efficient method well-suited for deployment on portable devices. We demonstrate the effectiveness of our approach on challenging examples involving rods, shells, and solids, showcasing its versatility and potential for widespread adoption.

Ball joints are commonly used in graphics and robotics in the simulation of articulated rigid body systems. The ball joint orientation can be decomposed, non-uniquely, into swing and twist components. In modeling with ball joints, it can be advantageous to include limit constraints on the swing, twist, or both. While many physics engines include such limit constraints, we have found that twist limit constraints, when present, are lacking in robustness, typically due to unhandled singularities in commonly employed swing-twist decompositions, where the twist state is undefined. In this paper, we introduce two novel, robust models of the ball joint twist limit constraint. The first utilizes the Euler angle decomposition of the orientation, which has two singularity points. To robustly handle the twist limit constraint, we propose a novel treatment at singularities based on the physical behavior of a three-hinge system. Thus, our model significantly reduces the dramatic motion that can occur around the singularities, eliminating the need to directly avoid them. Second, we introduce an alternative incremental model that avoids singularities altogether during simulations. In this model, the twist is quantified by integrating the instantaneous twist velocity in time, ensuring that the singularity shifts from one time step to the next and remains far from the current state. We investigate the properties of both proposed models. We demonstrate that both models offer robust and practical solutions to common issues encountered in constraining the motion of ball joints.

Recently, Zhao et al. [Zhao et al. 2024 ] proposed a new method for constructing signed distance functions from fluid simulation particles. This method was able to achieve superior surface smoothness, noise reduction, and temporal coherence compared with previous methods. One of the main limitations of the method was its relatively slow construction times, even though it utilized both the CPU and GPU. In this paper, we consider two modifications to this scheme that make the algorithm easier to optimize without introducing any perceptible changes in reconstruction quality, as illustrated in Figure 1 . With these improvements, a surface can be reconstructed from a single fluid simulation with 2M particles in 2.21 seconds, compared with 72.3 seconds for the original method, resulting in a single-frame reconstruction speedup of about 33 × , making the surface reconstruction fast enough for use within a simulation framework. When reconstructing surface for multiple simulation frames together, we achieve a speedup of about 5 × compared with the original. The optimized implementation will be released with the publication.

We address the problem of accurate capture of interactive behaviors between two people in daily scenarios. Most previous works either only consider one person or solely focus on conversational gestures of two people, assuming the body orientation and/or position of each actor are constant or barely change over each interaction. In contrast, we propose to simultaneously model two people’s activities, and target objective-driven, dynamic, and semantically consistent interactions which often span longer duration and cover bigger space. To this end, we capture a new multi-modal dataset dubbed InterAct, which is composed of 241 motion sequences where two people perform a realistic and coherent scenario for one minute or longer over a complete interaction. For each sequence, two actors are assigned different roles and emotion labels, and collaborate to finish one task or conduct a common interaction activity. The audios, body motions, and facial expressions of both persons are captured. InterAct contains diverse and complex motions of individuals and interesting and relatively long-term interaction patterns barely seen before. We also demonstrate a simple yet effective diffusion-based method that estimates interactive face expressions and body motions of two people from speech inputs. Our method regresses the body motions in a hierarchical manner, and we also propose a novel fine-tuning mechanism to improve the lip accuracy of facial expressions. To facilitate further research, the data and code will be made public.

Parametric human body models, especially from the family of SMPL, are widely used in fields such as motion capture, 3D human pose and shape estimation, as well as for animation or synthesis of human motion. A problem that arises when different body model types are to be used in the same workflow is that model parameters are not interchangeable. Existing methods to convert parameters are based on optimization. While these methods achieve low conversion errors, they exhibit long runtimes, making them suboptimal for the conversion of real-time data or large datasets. In this paper, we improve on these aspects by presenting a deep learning-based conversion approach that is both accurate and fast, enabling the interplay of different body model types in real-time scenarios. To this end, we use a fully-connected neural network to jointly convert translation, shape, and pose parameters for different SMPL-based body models and show that we can accurately convert sequences from different datasets. Our method is able to provide feedback on its conversion quality, which can be used to enhance the trustworthiness of a superordinate processing pipeline. We also implement a demo application for the real-time streaming and visualization of digital human twins. We use our conversion method to transform parameters of different body model types to a common representation before streaming and show that we can do so in real-time even for sequences that are running at 120 Hz.