Methodology: Personalised Physics
Our approach to personalise physics
Last updated
Our approach to personalise physics
Last updated
Our physics engine enables physics properties customisation via text (Fig. 10). We train on a corpus of five million words, decided based on the rule of 10, for semantic understanding and dynamic parameter adjustment [16]. We will pursue zero-shot 3D content creation solely with CLIP guidance.
Figure. 10: Text-Driven Physics Customisation Framework
Firstly, for dynamic parameter adjustment, the LLM employs semantic mapping that uses trained embeddings to map the tokenised text components to predefined physics parameters in the engine. Secondly, this is translated to actionable commands. Lastly, such commands will make dynamic parameter adjustment.
Figure 11: Example of Actionable Command Translated from Parsed Instructions
Figure 12: Physics Engine Features
At its core, the engine (Fig. 11) simulates both solid and deformable structures, using a NeRF-inspired finite elements dissection for detailed physical property analysis. It employs Finite Element Analysis (FEA) to anticipate structural responses to external forces, enabling rapid computation of stress and deformation under applied loads. Through text prompts, users can specify the load[1], so the engine will simulate and visualise the resulting stress distribution.
Optimised with octrees [13] and bounding volume hierarchies [11], the engine will integrate the Gilbert-Johnson-Keerthi (GJK) algorithm to pinpoint collisions [14] and use an impulse-based resolution system that determines appropriate collision responses. The engine also manages structural joints and constraints, ensuring authentic movement patterns of interconnected structures. Users can customise both collision properties[2] and joint management[3] via textual prompts.
