LLM Experiments with Programmable Matter: Preliminary Studies with SP-CellBots

This research note is a snapshot of an early phase of LLM-assisted interaction with programmable matter. It investigates how current general-purpose LLMs, without dedicated 3D training, can analyze modular bot clusters, devise spatial strategies, execute them, and assess the outcome via the SP-CellBots API. The note documents compact preliminary studies on replication, wall repair, morphology recognition, and functional Shepherd scenarios. The experiments show that modern models already provide usable responses to simple 3D and control tasks, especially when clear API feedback and structured prompts are available. At the same time, more complex spatial tasks remain prone to orientation errors, incomplete planning, and issues with connectivity or shape preservation. This repository contains the final PDF and source files for the research note.