AI game design in VR hunting

Abstract

In the scenario of a VR hunting game, behaviour and autonomy of in-game animals are crucial in minimizing users’ uncomfortableness and maximizing their experience while gaming. AI design of a successful Non-Player Character (NPC) mainly depends on four major behaviour layers which include: Decision, Steering, Navigation and Locomotion. In this study, the objective is to focus on designing and implementing novel algorithms for each of these four layers to achieve an integrated behaviour system in order to allow these animal characters to interact with the environment in an intelligent and realistic way. Specific and generic animal behaviour trees are designed and implemented on a typical predator animal with intelligent and cooperative consciousness in decision-making when interacting with a complex and ever-changing game environment. A novel physics-based rig simplification model: Constrained Ellipsoid Model (CEM) is proposed which allows steering inputs to generate force and torque driven movements. New adaptive Autonomous Animal Behaviour (AAB) algorithms are designed for animal agents to navigate around the environment, with a research focus on collective intelligence and tactical actions. Advanced strategies for a group of autonomous animals are developed in order to simulate a more realistic forest environment. Computational experiments and comparisons with animation results are presented, which show significant advantages over previous work. Forward and Backward Reaching Inverse Kinematics (FABRIK) is extended and used as the inverse kinematic solver to handle the locomotion of the autonomous animals. This Generic Quadruped Inverse Kinematic Model (GQIKM) presented is seamlessly combined with AAB algorithms so that the animals can perform all sorts of autonomous actions efficiently, supported with realistic and smooth animations. Finally, this inverse kinematic embedded motion control system is further improved by extending Proximal Policy Optimization (PPO) and designing specific reinforcement learning algorithms in order to bring animal agents’ level of intelligence further deeper down into the actual skeletal motion of bones and joints. All the algorithms in each behaviour layer presented in this research can be used individually or integrally for autonomous behaviour and motion control of other general, bipedal, quadrupedal and multipedal characters not only in computer graphics but also in fields such as vehicle navigation or robotic steering etc.