A Minimal Self-Perceiving Embodiment for Large Language Models

Abstract

Large language models occupy no physical space and cannot verify whether their outputs have landed in the physical world. We present a minimal hardware-software architecture that enables an LLM to perceive a shared environment, express within it, and receive the physical consequences of its own actions. The core mechanism is the self-perception loop: an output channel paired with a same-modality sensor, so that physical actions re-enter the model's observation stream as sensor measurements. The loop has two implementations: audio (buzzer ↔ microphone) and haptic (vibration motor ↔ accelerometer). Across 660 randomized trials, the audio loop confirmed emission on all 10 tones (+24 to +51 dB above ambient), and the haptic loop achieved 100% detection across all 11 effects. Beyond self-perception, the body senses direct human contact through a force-sensitive resistor and thermistor, and classifies gas sensor readings into olfactory categories in real time via a 207-parameter neural network (three classes, 92–94% accuracy). The body costs approximately €160 and is fully open-source. We frame this as relational embodiment: a physical presence designed for co-presence rather than task execution. The model can change; the body persists.