Neural Via Points for Muscle Wrapping
Abstract
Line-based musculoskeletal simulation is important in both computer graphics and biomechanics research. Software such as OpenSim can provide simulations of musculoskeletal dynamics and neural control where it would be difficult to measure with experiments. These tasks are done with the help of OpenSim’s various types of muscle paths, including straight muscles and via point muscles. However, there are limitations to these models, specifically when it comes to handling muscle mass. One common method used by these models is to combine the mass of the bones and soft tissues with the mass of the muscle, treating each body segment as a single body. This method has been shown to produce errors in inertia that can be quite large and variable. The main challenge in allowing more accurate muscle mass in musculoskeletal simulation is that the muscle mass points must be differentiable with respect to the joint angles. Adding mass for a simple muscle path represented as a single line is relatively easy. However, when the muscle path must go through multiple 3D “via” points, it becomes computationally more difficult. Furthermore, these via points may be functions of the joint angles, allowing them to become active or inactive based on the current joint configuration. We define these kinds of muscle paths as via point muscles. Our goal, therefore, is to generate a large training set involving many configurations of a via point muscle and to build a differentiable machine learning model from this dataset. To achieve this goal we first created separate models for a line muscle and a path point muscle. These models take in 3D points that include the muscle origin and insertion and percentage length alpha. The model for the path point also includes the 3D via point and auxiliary input such as the length of the muscle paths. These models output the corresponding 3D point along the muscle path. For the via point muscle, we take the model to 2D to simplify the computation. In our first attempt at this model, we allowed the inputs to be more randomized, which generalized the model’s capabilities. This allowed us to use the model for a wider variety of configurations. However, we also created a model where we simplified the inputs even further, and were able to see the best model performance from that, but its application only being for a smaller range of configurations. Because muscle simulations are typically done in 3D, we demonstrate a procedure where define muscle paths in 3D and transform them to the 2D bounds and transform the 2D outputs of the model back into 3D. For each of these models, we aimed for a loss within the 1e-6 range. With these results, we can use these models to predict muscle mass points for muscle calculations as a function of joint angle.
Subject
neuralneural networks
muscle
muscle paths
muscle wrapping
via points
OpenSim
computer graphics
computer science
musculosketal simulation
simulation
muscle mass
mass points
machine learning
Citation
Sun, Kyne (2023). Neural Via Points for Muscle Wrapping. Undergraduate Research Scholars Program. Available electronically from https : / /hdl .handle .net /1969 .1 /200264.
Related items
Showing items related by title, author, creator and subject.
-
Tuley, William Bayard (Texas A&M University, 1969)Not available
-
Sprowls, Robert Wayne (Texas A&M University. Libraries, 1974)Specific-pathogen-free (SPF) swine were given orally 15 mg/kg of chloroquine daily; 8 were sacrificed at 24 hours after 2-12 doses and 3 at 6, 10 and 16 days after completing 14 daily doses. Liver, myocardium and diaphragm ...
-
Lee, Nong-Shein (Texas A&M University. Libraries, 1977)Differences in adipose and skeletal muscle tissues between double-muscled (DM) cattle and normal cattle fed high energy or low energy rations were analyzed; 4 DM and 2 normal cattle were fed a high energy ration, while 3 ...