Enhancing Human Motion Simulation with Neural Networks

Abstract

Realistic simulation of human movement plays a crucial role across various domains, including medicine, robotics training, video games, and virtual reality. The primary challenge in these domains lies in generating physiologically plausible human movements. This challenge is typically addressed via trajectory optimisation, either in the muscle actuation-space – producing human-like movements but demanding computationally intensive simulations – or in the space of joint-actuation, reducing the complexity of the simulations but often generating movements that are non-feasible for humans to achieve. In a recent study conducted by Jiang et al., a method was introduced to convert a muscle actuation-based optimisation problem into an equivalent joint actuation-based problem. This transformation involves leveraging neural networks trained to approximate joint torque limits and energy functions. This work aims to reproduce the neural training procedure with a 2D model, and then use it in a problem of predictive gait simulation. In particular, the work focuses on the correction function R which allows the torque limit to be deduced. The initial objective involves grasping the neural network training procedure in order to adopt the methodology appropriately. This step is crucial in ensuring the training of adequate neural networks capable of accurately approximating the desired function. The second objective entails employing trained neural networks to address a predictive gait simulation problem. In this second part of the work, the results obtained using this method are compared with the original problem – a comprehensive model encompassing a skeletal structure and muscles. Additionally, a comparison is drawn with an alternative problem variant employing a simplified model, comprising solely a torque-driven skeletal structure.