Overview / Introduction
The ability to predict posture and motion realistically is the crux of any comprehensive effort to model humans. Consequently, we have developed extensive capabilities in these arenas, using optimization.
We are working towards providing the most comprehensive real-time posture prediction product available. Although methods have matured that simulate posture based on prerecorded data or animations, some of which involve solving multiple optimization problems for simulation refinement, we have developed a new approach to posture prediction. This approach is called direct human optimized posture prediction (D-HOPP). It affords the virtual human a substantial amount of predictive autonomy, enabling simulations that independently respond to infinitely many scenarios without any prerecorded data. It also provides a platform with which one can study how and why people move the way they do. With this approach, joint angles provide design variables for an optimization formulation that is solved only once for a predicted posture or motion. The problem is constrained primarily by requiring a specified end-effector (i.e. a fingertip, elbow, etc.) to contact a specified point, line, or plane. The end-effector positions in Cartesian space are determined from the joint angles by using the Denavit-Hartenberg (DH)-method, a robust and time-tested kinmetics technique stemming from the field of robotics. Joint limits are imposed as constraints and are based on anthropometric data. Skeletal dimensions are also based on anthropometric data, so skeletal and joint characteristics are easily modified. Human performance measures that represent physically significant quantities, such as energy, discomfort, etc., provide the objective functions. We contend that human posture and motion is task-based, meaning it is governed by different performance measures, depending on what task is being completed.
D-HOPP operates in real time. In fact, we have developed a new tool called Optimization-Based Inverse Kinematics (OBIK), which allows the user to manipulate and position avatars as desired. However, in stark contrast to other currently available tools, posture is automatically optimized with every frame.
With D-HOPP, incorporating additional capabilities is simply a matter of introducing new constraints and/or objective functions. For instance, we are able to dictate the orientation of different parts of the avatar, incorporate self-avoidance allowing the avatar to acknowledge his/her body, and incorporate multiple kinematic chains and end-effectors (previously a substantial challenge with robotics modeling). In fact, the user can specify any end-effector whether it is actually located on the body or not. In addition, the user can restrict such end-effectors to a specified point, bounded line, or bounded plane.
In order to govern the predicted posture, we are developing a suite of human performance measures, which currently includes joint displacement, effort, discomfort, change in potential energy, visual acuity, and visual displacement. In addition, we are developing new methods for combining various performance measures using multi-objective optimization (MOO). All of the performance measures can be evaluated at any point, whether or not they are used as an objective function with D-HOPP. This has lead to the development of zone-differentiation capabilities, where by the user can study not only an avatar’s reach envelope but also color contours indicating areas with particularly high performance measure values. This serves as a new valuable tool for ergonomic design.
Our approach to motion prediction is essentially an extension of D-HOPP. However, rather than using joint angles as design variables, conceptually we use curves of angle-versus-time as the design variables. These curves are represented as B-splines, and technically, the control points for the B-splines provide the actual design variables. We use constraints and objective functions similar to those with posture prediction, although they are evaluated at each time step. As with posture prediction, this approach provides a construct in which additional functionality is easily incorporated. For example, to consider dynamic problems and calculate torques at the joints, equations of motion and torque limits are used as additional constraints. In this way, one is able to conduct dynamic simulation and analysis without the usual cumbersome numerical integration.
Areas of research under this effort include the following:
- Real-time human simulation
- Real-time optimization
- Modeling of performance measures
- Human posture prediction
- Human motion prediction
- Self avoidance
- Multi-objective optimization
- Ergonomic analysis
Contact Info
https://www.ccad.uiowa.edu/vsr/contact
Related Publications
- Marler, R. T., and Arora, J. S. (2004), "Survey of Multi-Objective Optimization Methods for Engineering," Structural and Multidisciplinary Optimization, 26, 6, 369-395.
- Marler, R. T., and Arora, J. S. (2005), "Transformation Methods for Multi-objective Optimization", Engineering Optimization, 37, 6, 551-569.
- Marler, R. T., Yang, J., Arora, J. S., and Abdel-Malek, K. (2005), “Study of Bi-Criterion Upper Body Posture Prediction using Pareto Optimal Setsâ€, IASTED International Conference on Modeling, Simulation, and Optimization, August, Oranjestad, Aruba, International Association of Science and Technology for Development, Canada.
- Marler, R. T., Rahmatalla, S., Shanahan, M., and Abdel-Malek, K. (2005), "A New Discomfort Function for Optimization-Based Posture Prediction", SAE Human Modeling for Design and Engineering Conference, June, Iowa City, IA, Society of Automotive Engineers, Warrendale, PA.
- Farrell, K., Marler, R. T., and Abdel-Malek, K. (2005), "Modeling Dual-Arm Coordination for Posture: An Optimization-Based Approach", SAE Human Modeling for Design and Engineering Conference, June, Iowa City, IA, Society of Automotive Engineers, Warrendale, PA.