Automatic Scheduling for Parallel Forward Dynamics Computation of Open Kinematic Chains

Abstract

Abstract-Recent progress in the algorithm as well as the processor power have made the dynamics simulation of complex kinematic chains more realistic in various fields such as human motion simulation and molecular dynamics. The computation can be further accelerated by employing parallel processing on multiple processors. In fact, parallel processing environment is becoming more affordable thanks to recent release of multiplecore processors. Although several parallel algorithms for the forward dynamics computation have been proposed in literature, there still remains the problem of automatic scheduling, or load distribution, for handling arbitrary kinematic chains on a given parallel processing environment. In this paper, we propose a method for finding the schedule that minimizes the computation time. We test the method using three human character models with different complexities and show that parallel processing on two processors reduces the computation time by 35-36%.

Parallel Dynamics Computation Using Prefix Sum Operations

“…A similar simplification of the operand by isomorphism may be further investigated. After scanning the forward propagation (15), we proceed with the scan of the backward propagation (2) using the following linear recursion:…”

“…respectively. Therefore, the complete ABIA comprises two scan operations with operands A i 's given in (15), (16) for the computation of bias torques τ bias i 's, a nonlinear recursion (7) for the computation of the ABI, and then finally two scan operations with operands A i 's given in (18), (19) for the computation of joint accelerationsq i 's. We may also combine the two backward scans pertaining to (16) and (18) into a single backward scan, with the corresponding linear recursion given by:…”

“…At the outset, data-independent parallelism from different states and inputs suggests a parallel computation on the group level. Within individual groups, the datadependent parallelism in inverse and forward dynamics is well known to the robotics community, thanks to decades of extensive research on adapting recursive sequential algorithms [8], [9], [10] to parallel algorithms for multiprocessor systems [11], [12], [13], [14], [15]. Such algorithms are mainly intended for complex simulations with a large number (e.g., ≥ 1000) of links that are unusual for robotic applications.…”

Dynamical Simulation Priors for Human Motion Tracking

“…These models for the most part have only been developed and tested in the context of synthesis (i.e., animation [16], [19], [31], [33], [39], [60], [62], [61], [57]) and humanoid robotics [10], [28], [47], [59].…”

“…Dynamical simulation has a large body of existing work in animation [4], [16], [19], [21], [31], [33], [39], [60], [63] and robotics [10], [28], [47], [59] and is now a commodity technology. Simulation allows to computationally account for various physical and biomechanical factors that affect human motion, Fig.…”