01 · Form finding
Twist, taper, evaluate.
Galapagos searches twist, scale and count against a Ladybug radiation study.
Evolutionary solver exploring the form space
Form matrix
Candidates with their radiation roses
Reinforcement learning · UCL Digital Studio · with Raghul Siva · 2024
A twisted tower, panelled and rationalised — then a PPO agent learns to open and fold its skin from a week of simulated sun.
01 · Form finding
Galapagos searches twist, scale and count against a Ladybug radiation study.
Evolutionary solver exploring the form space
Form matrix
Candidates with their radiation roses
02 · Panelisation
Hexagons where the surface is gentle, triangles where it isn’t; relaxed with Kangaroo, clustered with a Gaussian mixture model.
Three towers and their unrolled grids
Relaxation — Kangaroo settling the mesh
Rules — 4 m slabs, ~1.3 m hexagon
GMM clusters mapped back
03 · Adaptive components
Folding experiments in Kangaroo. Each is a single-parameter action an agent can learn to drive.
01 Triangulated iris
02
03
04
05 Petal fold
06
07
08 Diamond fold
Three systems tracking the sun
Three systems, three towers — hexagonal scaling, triangular folding, diamond folding
Section — hexagonal scaling
Section — triangular folding
04 · Reinforcement learning
Observation: sun-vector radiation per panel. Action: scale factor. Reward: less radiation, less deployed surface. Trained with PPO on 1, 2, 4 and 8-day windows and tested on unseen days.
Training — one panel learning to close
Training — top view
Trained policy on the tower — apertures tracking the sun through a day
1 · 2 · 4 · 8 days of training
Hourly values — longer windows converge lower
RL policy vs parametric rule, same tower, same sun
Interactive
Aperture ∝ 1 − radiation. Move the sun around a curved tower face.