Adaptable Training Environments from a Single Capture: A 3D Gaussian Splatting Workflow for Mission-Relevant Scenario Variation
Adaptable Training Environments from a Single Capture: A 3D Gaussian Splatting Workflow for Mission-Relevant Scenario Variation
Virtual training environments often represent a single place under fixed conditions, although operational settings are variable, unpredictable and disruptive. Repeated training in an unchanged environment makes scenarios predictable and limits exposure to how reduced visibility, damage and changing conditions affect perception and decision-making. Yet producing variations of weather, lighting, season and physical disruption through repeated capture, manual modelling or physical staging is rarely practical.
Two recent developments enable a more flexible production workflow. 3D Gaussian Splatting (3DGS) reconstructs visually detailed scenes from ordinary imagery for real-time rendering, while generative AI can be used to modify source imagery to introduce conditions such as accumulated snow, surface wear or structural damage, keeping the underlying environment recognisable.
This paper presents a modular workflow that treats a 3DGS reconstruction as a reusable spatial foundation for multiple scenario variants. The source scene may be reconstructed from captured or AI-generated imagery, allowing the study to examine transferability between the two. The key value is repeatability: variants of one scene can serve different mission-relevant tasks, such as route familiarisation in daylight, reconnaissance at night, or casualty evacuation and damage assessment after collapse and snowfall.
Transient conditions, such as night-time lighting, fog, smoke, or rainfall, can be introduced at runtime without modifying the underlying asset. Persistent changes, such as accumulated snow, worn surfaces, debris and structural damage, require source-image modification and reconstruction of a new 3DGS asset. Hybrid conditions combine both.
The variants support a hypothesised progression in which complexity increases with the trainee's familiarity and the demands of the task. A new environment is first presented under full lighting with a clear representation. As familiarity grows, reduced lighting, night-time conditions, approximated infrared or ultraviolet illumination, seasonal change and clutter demanding material recognition are introduced. For rapid mission preparation, the principle is just in time, just enough: the minimum sufficient fidelity for the task at hand.
The study evaluates whether editing a limited subset of source images produces sufficiently coherent persistent changes despite inconsistencies and hallucinations introduced by generative AI. Approaches are assessed for visual coherence, cross-view consistency, spatial alignment with the baseline, production effort, transferability between captured and generated environments, and suitability for specific tasks and progression stages. The emphasis is on sufficient visual fidelity rather than physical or sensor-level accuracy.
During preliminary tests, five variants of one base scene were produced. Each variant was generated and integrated as navigable 3DGS assets into a virtual training application within 35 minutes in total. The baseline scene remained recognisable and navigable.
The intended outcome is a practical decision framework identifying which environmental components are best introduced at runtime, before reconstruction or through a hybrid approach, and which level of complexity suits each training stage. A single capture can thus yield a library of mission-relevant variants, supporting less predictable, more operationally relevant immersive training that evolves with operational requirements and emerging threats.