USING HIGHTHROUGHPUT COMPUTING FOR DYNAMIC SIMULATION OF BIPEDAL - PowerPoint PPT Presentation

USING HIGH–THROUGHPUT COMPUTING FOR DYNAMIC SIMULATION OF BIPEDAL WALKING Mohammadhossein Saadatzi, Anne K. Silverman, Ozkan Celik

Outline  Simulation framework for combined human bipedal walking and wearable robots  Introduction  Problem definition  Using HTC to solve the problem  Applications of the prepared framework  A passive hip exoskeleton  Walking energetics at different speeds 1

Introduction  Previously, high-throughput computing (HTC) was used for stochastic simulation and analysis of human movements [1,2].  By considering parametric uncertainty in musculoskeletal model, HTC was used to investigate cartilage loading in the knee during movement.  Simulations of movement are traditionally performed using generic models with assumed parameters and geometries.  Population variability and parametric uncertainty become critically important to consider when interpreting model predictions [1,2]. [1] Smith CR, et al. J Knee Surg, 29 (2), 99-106, 2016. 2 [2] Smith CR, et al. J Biomech Eng, 138 (2), 021017, 2016.

Introduction  Simulations and dynamic optimization for investigation human biomechanics and human-robot interaction.  We have developed a framework for combined simulation of human bipedal walking and exoskeletons [1].  Simulation of human bipedal walking through muscle actuation presents an optimization problem with a substantial number of parameters [2,3]. [1] Saadatzi M, et al. Proceedings of 41 st ASB, 451–452, 2017. 3 [2] Miller RH. J Biomech, 47 (6), 373–1381, 2014. [3] Umberger BR, J. Royal Soc. Interface, 7 (50), 1329-1340, 2010.

Introduction  Calculation of the cost function for this optimization problem, necessitates significant computational power [1].  Because of the redundant nature of the human musculoskeletal system, the problem is highly prone to getting stuck in local minima.  Researchers have chosen to solve the problem several times [2-5].  Miller [2], single step, 4 times, 2.5 days using HPC.  Umberger [3], single step, 10 times.  Ong et al. [4], 10 seconds, 20 times.  They used the best solution to seed another round of 20 optimizations.  This process was repeated until the change was less than 5%. [1] Ackermann M, van den Bogert AJ, J Biomech, 43 (6), 1055–1060, 2010. [2] Miller RH. J Biomech, 47 (6), 373–1381, 2014. [3] Umberger BR, J. Royal Soc. Interface, 7 (50), 1329-1340, 2010. 4 [4] Ong CF, et al. IEEE Trans Biomed Eng, 63 (5), 894–903, 2016. [5] Ong CF, et al. Proceedings of ISB-TGCS, 19–20, 2017.

Introduction We use HTC as a tool to perform computations related to predictive simulation of human bipedal walking through muscle actuation. 5

Problem definition Actuation profile Human-exoskeleton combined model Forward dynamic simulation Exoskeleton structure and control parameters Human-exoskeleton kinematic and dynamic data Objective Optimization function Algorithm [1] Farris RJ, et al. IEEE Trans Neural Syst Rehabil Eng, 19(6), 652-659, 2011. 6 [2] Delp SL, et al. IEEE Trans Biomed Eng, 54 (11), 1940-1950, 2007.

Problem definition: Human-exoskeleton model The musculoskeletal model has 9 sagittal degrees of freedom and 18 muscle groups (24 muscles):  DOFs: pelvis tilt, two planar pelvis translations, hip and knee flexion/extension, ankle plantar/dorsiflexion  Muscle groups: soleus, vasti, gastrocnemius, tibialis anterior, bicep femoris short head, hamstring, rectus femoris, gluteus maximus, and iliopsoas. 7

Problem definition Actuation profile Human-exoskeleton combined model Forward dynamic simulation Exoskeleton structure and control parameters Human-exoskeleton kinematic and dynamic data Objective Optimization function Algorithm [1] Farris RJ, et al. IEEE Trans Neural Syst Rehabil Eng, 19(6), 652-659, 2011. 8 [2] Delp SL, et al. IEEE Trans Biomed Eng, 54 (11), 1940-1950, 2007.

Problem definition: Forward dynamic simulation Two controller states are considered, corresponding to stance and swing phases. Heel strike Toe off Heel strike Stance Swing 9

Problem definition Actuation profile Human-exoskeleton combined model Forward dynamic simulation Exoskeleton structure and control parameters Human-exoskeleton kinematic and dynamic data Objective Optimization function Algorithm [1] Farris RJ, et al. IEEE Trans Neural Syst Rehabil Eng, 19(6), 652-659, 2011. 10 [2] Delp SL, et al. IEEE Trans Biomed Eng, 54 (11), 1940-1950, 2007.

Problem definition: Actuation/exoskeleton parameters Actuation profile of each muscle is defined as two piecewise linear functions  One with 6 parameters for stance phase, Exoskeleton and structure and control parameters  One with 4 parameters for swing phase. The values of signals at discretized points constitute the optimization parameters. 11

Problem definition Actuation profile Human-exoskeleton combined model Forward dynamic simulation Exoskeleton structure and control parameters Human-exoskeleton kinematic and dynamic data Objective Optimization function Algorithm [1] Farris RJ, et al. IEEE Trans Neural Syst Rehabil Eng, 19(6), 652-659, 2011. 12 [2] Delp SL, et al. IEEE Trans Biomed Eng, 54 (11), 1940-1950, 2007.

Problem definition: Objective function  A weighted combination of several cost functions constitutes our optimization Objective function goal:  Duration of walking  Target velocity error  Excessive alternating motion of the upper body  Metabolic expenditure of muscles 13

Problem definition Actuation profile Human-exoskeleton combined model Forward dynamic simulation Exoskeleton structure and control parameters Human-exoskeleton kinematic and dynamic data Objective Optimization function Algorithm [1] Farris RJ, et al. IEEE Trans Neural Syst Rehabil Eng, 19(6), 652-659, 2011. 14 [2] Delp SL, et al. IEEE Trans Biomed Eng, 54 (11), 1940-1950, 2007.

Problem definition: Optimization Algorithm  Covariance Matrix Adaptation (CMA) [1]. Optimization Algorithm  Previously has successfully been used for simulation of bipedal walking [2-4].  Available with OpenSim API.  We use the computed muscle control (CMC) method to generate favorable initial optimization parameters,  which reduces convergence time and probability of getting stuck in local minima. [1] Hansen N, et al. Evol Comput, 11(1), 1–18, 2003. [2] Wang JM, et al. ACM Trans Graph, 31(4), 2012. 15 [3] Dorn TW, et al. PLOS ONE, 10 (4), 1-16, 2015. [4] Ong CF, et al. Proceedings of ISB-TGCS, 19–20, 2017.

Methods  HTC enables the use of many computers and cores simultaneously HTC  Ideally, we would like to combine the results from several computers to reduce  necessary time for the convergence of our optimization  probability of getting stuck in local minima [1] Pordes, R et al. J Phys Conf Ser 78 (1), 012057, 2007. 16 [2] Sfiligoi I, et al. IEEE-CSIE, 428–432, 2009.

Methods: High-level algorithm  We run several instances of our optimization, and combine their results periodically to generate a more efficient optimization algorithm.  We seed a round of optimization on remote computers with different sets of initial optimization parameters.  We compare and combine their results to generate initial optimization parameters for subsequent optimizations. Remote computers Submit server Compare & Convergence combination Detection Initialization Optimization Optimization Optimization 17

Methods: High-level algorithm  Directed Acyclic Graph Manager (DAGMan): a meta- scheduler for the execution of programs (computations)  DAG file is used by DAGMan when submitting programs. PS.dag Prescript # Predictive Simulation DAG InitPop.sh Job PSitr PSitr.sub PSitr: PSitr.sub Job SCRIPT PRE PSitr initPop.sh J 1 J 2 J 3 J 4 J n SCRIPT POST PSitr nextItrPrep.sh RETRY PSitr 10 NextItrPrep.sh Postscript Retry 18

Methods: High-level algorithm  Prescript bash “initPop.sh” generates the initial population.  Postscript bash Prescript InitPop.sh “nextItrPrep.sh” generates the population for the next PSitr: PSitr.sub Job iterations. J 1 J 2 J 3 J n J 4  We used the best results to seed the following iterations of Postscript NextItrPrep.sh optimizations. Retry  Job “PSitr”, runs the sub- jobs, which consist of the optimizations of each round. 19

Methods: Individual jobs PSitr.sub # Jobs submit file executable = myExe.sh arguments = $(Process) transfer_input_files = bipedWalkingOptFiles.tar.gz, pop/bestParametersBinary.$(Process) transfer_output_files = bestSoFarLog_$(Process).txt, bestParametersBinary_$(Process) periodic_remove = (CurrentTime - QDate) > 4200 queue 100 20

Results & discussion  We tested different intervals of running individual jobs, and explored fixed and adaptive initial step-size of the CMA algorithm Doesn’t walk the entire duration OSG 8-core computers, and AWS 36-core computer Walks the entire duration, when gets smaller becomes more realistic  OSG 8-core computers (3.89) and AWS 36-core instance (3.86). 21