Jumping is useful as a precursor for flight and arboreal locomotion, as well as locomotion through intermittent terrain. This thesis aims to answer the questions: What insights can be gained through the application of predictive modelling methods to novel contexts in jumping dynamics problems, and how should the fitness of different predictive modelling techniques be assessed? This thesis uses the metrics: predictive power, computational cost, and intellectual investment, to assess predictive modelling methods for use in jumping research. Analysis of model equations is rarely found applied to multiple degree of freedom jumping models in the literature. This work explores such applications of analysis and determines that the method is capable of providing insights into low and high complexity models. The work identifies static optimisation for the resolution of second order kinematic redundancy and reinforcement learning as predictive modelling methods which are not currently used in jumping research. The results of this work imply that static optimisation could be made viable for providing insights into jumping dynamics problems. However, the static optimisation method is demonstrated to produce results which are comparable to experimentally measured data found in the literature and may be a useful method in jumping applications. The method is limited in that it does not accommodate the physical constraints of jumping systems. Deep Deterministic Policy Gradients, a reinforcement learning method, is shown to also violate physical constraints. The method is demonstrated to be unreliable when applied to jumping dynamics problems with 32% of optimisation runs failing to achieve any positive rewards. Dynamic optimisation is assessed on its predictive power, computational cost, and intellectual investment and is deemed to require considerable intellectual investment to use. As well as insights into the methods used to solve jumping problems, this work provides technical contributions of: 1) demonstration of a novel method used to determine the dynamic balancing capability of a jumping system and 2) empirical demonstration of a second order kinematic simulation method with results comparable to those found in the literature.
|Date of Award||31 Dec 2022|
- The University of Manchester
|Supervisor||william crowther (Supervisor) & Ben Parslew (Supervisor)|
- Reinforcement Learning
- Predictive Modelling