SymbioShell operates as a closed loop of sensing, cognition, and actuation, running entirely on edge hardware with no cloud dependencies.

The biological subject, Physarum polycephalum, is housed inside a cylindrical chamber containing a Petri dish with an agar-and-oats substrate. A Raspberry Pi Camera Module mounted on the chamber lid captures macroscopic images of the organism every twelve minutes. Image acquisition is synchronized with arrays of infrared LEDs that briefly illuminate the organism only during capture and switch off immediately afterward, minimizing photic stress on the photosensitive slime mold.

Each captured frame is downsampled to 128×128 px and processed by a lightweight convolutional autoencoder. The encoder compresses each image into a 32-dimensional latent vector, an internal geometry free from anthropocentric labels, designed to encode morphological and topological features the organism produces over time. Training is performed offline in Python using TensorFlow/Keras over an initial corpus of observations. The trained encoder is exported to TensorFlow Lite and runs locally on a Raspberry Pi 5.

Latent vectors are clustered using k-means via scikit-learn, generating an unsupervised taxonomy of biological states. During operation, each new latent vector is compared against stored centroids. If its Euclidean distance exceeds a calibrated threshold, the system registers a novel behavioral state. Persistent memory of states, gestures, and timestamps is maintained in JSON files (gesture_map.json), allowing the vocabulary to grow incrementally without retraining.

Environmental conditions are captured by an SHT3x temperature/humidity sensor communicating with the Raspberry Pi via I²C. Each visual capture is paired with an environmental reading and logged to CSV alongside timestamp, active LED group, and image filename, building a longitudinal dataset of the biosystem.

Each detected state is mapped to a discrete gesture. When a state is novel, a generative module improvises a finger-and-wrist configuration constrained by the mechanical safety envelope of the prosthesis, and indexes it to that state for future reuse. The robotic prosthetic hand has six degrees of freedom: one per finger plus the wrist. It is driven by servomotors and built from 3D-printed components. Motion commands travel via serial communication from the Raspberry Pi 5 to an Arduino Mega, which handles low-level servo control. The wrist range is mechanically restricted to prevent damage at joint extremes.

A live interface, rendered on an external display, exposes the AI's reasoning to the audience as a real-time stream of system states: capture of the current organism state, visual feature analysis, perceptual synthesis between biological dynamics and microhabitat, latent-space mapping, autonomous clustering into non-human logical categories, validation of biological agency, and indexing of syntax and motor articulation of each gesture. This visible cognition reveals the otherwise opaque decision loop between organism and machine, while preserving the fact that the system operates without human supervision.

The full stack, including acquisition, edge inference, clustering, gesture synthesis, environmental logging, and motor articulation, runs locally on the installation as a self-contained ecosystem of perception, cognition, and articulation.