Animating Virtual Characters using
Thomas Geijtenbeek, PhD Thesis
A Quest for Realistic Animation
Believability is essential for virtual environments; any anomaly that can distort the acceptance of an alternate reality must be avoided. In cases when virtual environments attempt to imitate our world, a disrespect of our laws of physics is a common source of aberration. The incorporation of physical constraints offers an ongoing challenge for the animation of characters and objects that inhabit such a virtual environment.
Traditionally, the motion trajectories of virtual characters and objects are hand-crafted by skilled animators. This is a time-consuming process, and the resulting animations are purely kinematic: they have no regard for force or mass. The physical validity of such motion trajectories depends solely on the skill of the animator. As an alternative, people have been capturing performances from live actors. The resulting trajectories are guaranteed to describe natural motion, but the approach is limited to characters and motions for which an equivalent live performance is available. Furthermore, the naturalness of these motions may not be preserved during the processing that is often required to leverage limited motion data.
Animation becomes even more challenging during interaction; there are many ways in which virtual entities can interact with each other, and subtle variations may call for substantially different responses. For example, a stack of boxes may collapse in infinitely many ways, even with similar initial perturbations. Animation systems that draw from existing data (either captured or hand-crafted) require a complex process involving events, rules and geometric processing to generate proper responses. Despite great advances in data-driven methods in the last decade, their ability to produce plausible and non-repetitive responsive animation is restricted by the contents of the database.
The Power of Simulation
Physics-based simulation offers a fundamentally different approach to computer animation. Instead of directly manipulating the motion trajectories of objects and characters, this approach lets all motion be the result of a physics-based simulation process. As a consequence, physics-based characters and objects automatically interact in accordance with the laws physics, without the need for additional motion data or scripting. Over the past decades, physics-based simulation has become an established method for the animation of passive phenomena, such as cloth, water and rag-doll characters. The conception that physics-based simulation can also be used to simulate actively controlled characters dates back to the early stages of 3D computer animation, and has incited many research papers since. However, commercial animation frameworks still resort to kinematics-based approaches when it comes to animating active virtual characters.
To understand this reservation, it is important to recognize the scope and complexity of controlling simulated characters. Just like with real-world entities, the pose of a physics-based character is controlled indirectly, through forces and torques generated by actuators that reside inside the body. As a result, the global position and orientation of a physics-based character cannot be controlled directly, but only through deliberate manipulation of external contacts. This impediment -- also referred as underactuation -- poses a direct challenge to basic tasks such as balance and locomotion. Such a challenge has no equivalent in traditional kinematics-based animation; it bears a much closer relation to humanoid robotics and control theory.
On the other hand, the primary focus of animation is visual quality. Several (especially early) physics-based controllers that emerged from robotics research focus on robustness, and have little attention paid to motion style. Consequently, many of these otherwise impressive results have been deemed stiff and robotic when compared to data-driven alternatives. An important insight here is that even though these simulated characters move in a way that is physically valid, their motion is not necessarily biomechanically accurate. The amount of torque that can be produced by real-world muscles is restricted by various physiological and mechanical properties, and biological control systems possess significant delays in neural processing and muscle activation. These biomechanical constraints are incorporated in any behavior we witness in nature, and play an important role in the perception of naturalness. However, the actuation models used in physics-based character animation are often highly simplified and do not enforce any of these constraints. As a result, such simulated characters often behave differently from biological equivalents, even if their motion is optimal within the constraints of the model.
A Need for Biomechanical Constraints
One possible way to overcome this is to incorporate captured motion data into the control strategy, but this limits motions to be similar to the data that is available. More recently, researchers have begun incorporating biomechanical constraints into the character model, based on results from biomechanics research. This trend is gradually changing the direction of physics-based character animation research, and is further broadening its scope.
To implement all these aspects into a robust and flexible framework is a daunting task. Even though professional physics-based simulation software has become readily available, successful implementation of a physics-based character animation framework requires at least some knowledge of multi-body dynamics, numerical integration, control theory, biomechanical modeling and optimization theory. In addition, many physics-based control strategies require skillful and time-consuming manual tuning before they can be put to use -- a process often poorly documented in research publications. Finally, physics-based character animation is computationally much more expensive than kinematics-based alternatives. It is only since about a decade that passive physics-based characters can be simulated in real-time on a consumer-grade PC. Currently, real-time performance of many state-of-the-art controllers is limited to a single character at a time on modern hardware.
The Future of Physics-Based Characters
In spite of these ramifications, recent trends have shown a renewed interest in using simulated physics for interactive character animation (see the Figure above). After decades of floundering, the field is maturing, with many recent publications demonstrating tremendous progress in both robustness and visual quality. As such, physics-based character animation remains an exciting research topic that is likely to play an increasingly important role in computer animation in the years to come.
For more information: t.geijtenbeek [at] uu [dot] nl