Virtual reality (VR) is increasingly being used in companies to enhance the workflows of industrial domain experts. However, the effective use of VR and virtual spaces relies on 3D content, including models and animations typically created by graphic artists. This PhD thesis aims to streamline the industrial application of VR by addressing the challenge of enabling industrial experts to create 3D animations. We approach this from two areas of research: immersive object animation, and exploded view animations.

Within the field of object animation, we developed and evaluated a VR animation tool, specifically designed for industrial end users. The tool uses performance animation to record the motion of objects, visualized by a motion path in 3D space. This natural approach simplifies the animation process, allowing non-animators to create animations efficiently without any prior experience, demonstrated by a mixed methods user study involving industrial users. The motion path approach was extended for multi-object animation, again focusing on ease-of-use for novice animators. To control the timing of objects, a timeline was introduced with layers for each animated object. The tool was compared with the animation system from a popular VR application, highlighting the superior usability as well as areas of improvement. These tools represent a step towards streamlining industrial animation processes, making animation more accessible to industrial VR users, and promoting a more efficient use of VR in industrial settings.

Within automatically generated exploded view animations, we explored the use of traditional assembly sequence planning methods in a VR tool, demonstrating real-time computation through simplification and abstraction. We then presented the world’s first machine learning-based approach, combining a traditional assembly-by-disassembly process with deep learning methods. We first introduced a novel point cloud dataset, AssemblyNet, for the multi-class classification of disassembly directions of parts. The dataset was benchmarked using four well established methods as well as two variations of a novel two-path network architecture. Building on this, we introduced a second dataset for the binary classification of whether a part can be disassembled or whether it is blocked. We also improved our two-path network approach, presenting the Point-based Disassembly Attention Network (PointDAN), incorporating a novel distance-based attention mechanism. We integrated the deep learning methods into our VR tool and demonstrated its accuracy and the potential industrial impact through a quantitative analysis and an expert user study.