High-performance data-driven control of physically-based human characters

This thesis describes a system to enable responsive user guided control of a physically simulated human character.The control system is meant to be robust to disturbances while also producing movements that are of a similar quality to the visuals produced by more high quality kinematic animation systems used in modern video games.This task is difficult because human characters must dynamically retain balance through contacts with an environment, and walking requires control of an underactuated dynamic system.Simulation also does not guarantee a character will move in a natural manner, so care has to be taken to ensure visually unusual behaviours do not occur as a result of control.Work in the field of reinforcement learning has demonstrated the possibility of generating physical character control policies that imitate human motions with a high degree of success.Many methods have focused selectively on generating controllers that produce high quality motion, while important factors such as responsiveness, user controllability, motion diversity, and runtime costs have been somewhat overlooked.The approach presented here focuses on improving performance with respect to all these factors.A data-driven kinematic character controller sequences and blends motion capture data in order to generate medium-term kinematic motion plans which fit user controlled high-level goals.This allows movement direction, heading direction, speed, and style of motion to be responsively altered in a real-time user controlled manner, while also capturing subtleties of human behaviour in the data.Reinforcement learning is then used to train a simulated character controller that is capable of imitating the motion of the kinematic character controlled by a user.This necessitates a training scheme that captures the full distribution of behaviours that a human is likely to use, and which enforces the learned behaviour to retain the stylistic characteristics of the generated motion while making it physically feasible.The design of this system is also made with runtime cost in mind, ensuring that the result is useful in the context of real world application in video games where performance budgets are strict.i This project would not have been possible without the support, opportunities, and advice provided by Professor James Richard Forbes.I am very grateful for his supervision and guidance throughout the

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.