TransformerMorph: Local Transformer Policies for 3D Morphing of Modular Robots – A Feasibility Study

This technical note describes TransformerMorph v001, an experimental approach to learning-based 3D morphing in SP-CellBots. A compact Transformer policy selects discrete vehicle-kinematics actions from a local 5×5×5 EGO view, a goal description, and a short action history. The behaviour was not trained by imitating predefined paths; instead, it was learned incrementally through Q-learning in simulated miniature worlds. The current implementation operates sequentially and without a global map. In zero-shot tests on 33 unknown test objects, the greedily executed policy completed 14 full morphs; with limited exploration, it completed 17. The results are not intended as a competitor to A*, but as a feasibility study for locally observing learned morphing policies. The note discusses the architecture, the training curriculum, typical failure patterns, limitations, and hybrid augmentation through an A* fallback.