Lux Automata: 635,040 LEDs with Homeostatic Drives

Physical Sculpture Architecture

Created through a three-year R&D collaboration (2019–2022) between Jason Bruges Studio and Google DeepMind, Lux Automata occupies the main lobby ceiling of DeepMind's King's Cross headquarters. The physical matrix consists of 42 × 10 × 7 voxels—2,940 custom cellular black-PCB cubes suspended on 6mm brass rods.

Each cellular cube houses 216 individually addressable LEDs at a 3mm pitch, totaling 635,040 variable-brightness LEDs. This volumetric arrangement creates three-dimensional light forms visible from the ground floor reception, upper mezzanine walkways, and the street exterior.

Six ambient foyer motion-detection cameras and a dedicated interactive pose/motion camera convert human movement in the lobby into low-resolution activity blobs (with zero video recording or storage, preserving complete privacy) that are injected directly as live cells into the 3D simulation substrate.

Four Families of Generative & Artificial-Life Substrates

1. 3D Rule-Based Cellular Automata (Discrete & Continuous)

Executed in real time in Unity, the sculpture runs both hand-crafted and evolutionarily discovered 3D CA rules—producing volumetric gliders, oscillators, solitons, "pulse engines," "fizzy rings," and weather-like "glider clouds" that collide and react when human movement injects live cells from below.

2. Neural & Developmental Cellular Automata ("Soft CAs")

Python-based continuous cellular automata where each voxel's next state is governed either by an evolved 3D Convolutional Neural Network (CNN) or by a Developmental Gene Regulatory Network (GRN). In the GRN mode, cells communicate via diffusing virtual transcription factors and visualize gene expression patterns inspired by biological embryogenesis—a structural nod to AlphaFold.

3. Reinforcement Learning Agents in CA Worlds

Single-agent and multi-agent RL policies (DQN and self-play) operate inside the 3D cellular automaton—learning to stamp 4×4 cell patterns to control wave propagation in real time, or competing against one another in emergent volumetric games such as 3D CA Pong (a nod to DeepMind's Atari heritage).

4. Artificial-Life Population Simulations

Multi-agent ecological simulations of evolving populations—exhibiting foraging, predator-prey pursuit and evasion, and flocking/congregation—rendered directly within the 3D voxel matrix.

The Embodied AI Curator & Homeostatic Drives ("Agatha")

At the heart of Lux Automata is a drive-based action-selection agent ("Agatha") that curates which generative rules and interactive modes run at any moment:

• Internal Homeostatic Drives: The agent balances competing internal needs (ranging from 0.0 to 1.0) including Novelty, Excitement, Calmness, Familiarity, Human Interaction, and Foyer Influence—seeking equilibrium between staccato, high-energy bursts and calm, meditative states.
• Computational Aesthetic Appraisals: During self-play and evolutionary search, the agent evaluates 3D patterns using mathematical aesthetic metrics—including Birkhoff's Aesthetic Measure (M = Order / Complexity), scale-invariant fractal dimension, Kolmogorov compressibility, block entropy, glider mobility counts, and Mutual Information Neural Estimation (MINE) for human engagement.
• Minimal Turing Play & Interactive Mode: When an individual stops beneath the interactive camera, the sculpture can transition from ambient curation into dyadic play—engaging in non-verbal turn-taking, movement mirroring/imitation with variation, and simple RL chase/evade games.