Chemical accidents pose substantial risks to personnel, surrounding communities, and the environment, underscoring the importance of effective safety education and training. Virtual reality (VR) provides immersive and risk-free simulations, yet many systems still prioritize realism, which may not always effectively support hazard perception. This study investigates how different accident presentations—Realistic versus Exaggerated—affect learners' perception, emotional responses, and safety-related learning in a $2 \times 3$ mixed design across Spill, Fire, and Explosion scenarios. Results show that Exaggerated presentations enhance arousal, perceived risk, and safety awareness, while Realistic presentations support immersion and understanding of accident mechanisms. The three accident scenarios elicited distinct response patterns. These findings offer practical guidance for designing VR-based safety training and suggest broader applications for immersive educational environments.
Immersive Analytics (IA) utilises stereoscopic 3D (S3D) graph visualisations in Virtual and Augmented Reality to harness spatial understanding and engagement beyond 2D screen displays. Despite growing research interest, little is known about how users perceive such structures, or how their gaze behaviour relates to task performance. We address this through a series of eye tracking studies on visual analytical tasks over S3D graphs in VR. Our first study, on single-layer graphs, reveals systematic links between gaze behaviour and task performance, with distinct exploration strategies across tasks. We then extend to multilayer graphs and show how their spatial arrangement and complexity reshape gaze behaviour. Finally, by mining the gaze patterns of successful analysts, we enhance the visualisations with targeted visual cues; in a follow-up study, these improve task correctness in the majority of trials. Overall, our findings deepen the understanding of graph perception in immersive environments, from single to complex multilayer structures, and demonstrate how eye tracking can both reveal reading strategies and inform S3D graph design through gaze-derived cues.
Passivity constraints bound the impedance range that force-feedback devices can stably display, so real-time haptic training defaults to linear physics—typically a mass–spring–damper (MSD) model—although materials worth training on are nonlinear and non-Newtonian. We define this discrepancy as the rheology gap and ask which motor-skill components survive a deliberately linearized rendering of an artisanal dough-stretching task. A within-instance design held the task fixed while varying VR factors across five conditions ($N=378$); real-world transfer was assessed in a separate cohort ($n=90$) evaluated on authentic dough by blinded certified judges (AVPN-derived rubric), with published rheometry as the material force reference. Relative to video instruction, VR training produced a medium real-world benefit (Cohen's $d=0.76$): procedural and spatial components transferred, whereas material-specific force modulation did not—thickness uniformity ($-$41.6%) and process efficiency ($-$45.6%), the most rheology-sensitive components, showed the largest deficits. Non-informative force feedback impaired learning, and gains persisted at 24 h. Because conditions covary on several factors, effects are reported as composite. Our central contribution is a reproducible, component-level benchmark—task protocol, objective metrics, characterized linear baseline, and published-rheometry reference—on which any future nonlinear engine (FEM, MPM, XPBD, neural physics) can be scored under identical conditions, rendering simulation fidelity falsifiable. As a first application, a simulation-stage prototype rendering the calibrated strain-stiffening law under time-domain passivity control runs stably at 1 kHz and cuts force-fidelity error from 41.0% to 24.6% NRMSE. Pending closed-loop validation of the strain-stiffening hypothesis (offline $R^{2}=0.31$), current VR is best positioned as pre-apprenticeship scaffolding, not a substitute for physical practice.
The generalized winding number (GWN) offers a robust solution for point membership classification (PMC), tolerating geometric imperfections such as non-watertight boundaries. While GWN evaluation on discrete representations (e.g., meshes) has demonstrated its efficiency, its computation on B-Reps with trimmed surfaces remains computationally expensive, hindering practical use. We present a fast GWN evaluation framework for B-Reps with bounded error. The key idea is to adaptively discretize boundary curves with explicit analysis on deviation bounds. With the consistent discrete representations, queries inside and outside the surface bounding boxes are distinguished, and errors on the shared curves of adjacent surfaces are canceled out. Queries outside surface bounding boxes are reduced to efficient mesh-based evaluations with bounded error, which further enables efficient spatial acceleration. In addition, an optimized spherical projection that is both fast and accurate is designed for the remaining queries in boxes. Numerical experiments demonstrate the approach's superior efficiency while avoiding misclassification, compared with SOTA GWN approaches, and classic open-source CAD kernel Open CASCADE.
Accurate modeling of 3D hand-object interactions plays a crucial role in modern 3D animation, serving as an essential component in animated movies and video game development. Traditional workflows require artists to manually adjust hand joints to create hand-object interactions, which is labor intensive and inefficient. We propose a novel framework for modeling 3D hand-object interactions from free-hand sketches. Our method generates physically plausible hand poses while ensuring precise alignment with user inputs. To achieve this, we introduce a pipeline that attentively integrates 2D sketch and 3D object information to model dexterous hand interactions. The core design principle is effective fusion of 2D and 3D features with explicit awareness of hand-object contact relationships. The overall framework mainly consists of three components, including a voxel-based feature fusion module for multi-modal representation, a hierarchical attentive prediction module for modeling long-range dependencies while capturing fine-grained details and a graph-based contact-aware adjustment module for refinement. Our method achieves strong performance and generalizes well to novel object categories. These components together yield a robust, user-friendly sketch-based 3D hand object interaction modeling system. Extensive experiments have verified the superiority of our proposed method compared with state-of-the-art works.
