HEAPS · SEVEN OF A KIND UNDER A CIRCLING SUN · HOW HIGH, HOW MANY ON OTHERS
THE BEST NOW · BY MERIT (SCORE × SIMPLICITY)
MERIT OVER TIME
Merit = score × simplicity. Best: the best robot so far. Mean: the average over all kinds on the map. Ticks: the epochs.
[?] WHAT HAPPENS HERE
On a Mac mini in a studio, solar BEAM robots evolve day and night: each one a circuit board — or two or three slotted together — with solar cells, a Tilden solar engine and a motor, gear motor or solenoid, and some wire. Every robot is built and tried in a physics simulation with a twin: dropped onto it (does it stay on top, does it interlock?), side by side under a lamp (does it climb or jump onto the other? how far does it lift itself on the other, how long do they touch?), alone (how far does it get?). The robots that do well with their twin are then tried as a group: seven of a kind, side by side under a sun that circles once — do they get up on each other, rest there, and stay? A robot shaded by others goes still, one on top stays in the sun and keeps moving, as in the towers of fire ants. The heap counts most. The map keeps the best robot of each kind — how it moves × how many parts it has (MAP-Elites, Mouret & Clune 2015) — and the evolution mutates and crosses them. Every two hours, when something improved, an epoch is published here. Select any robot to see it live, read its drawing and parts list, follow its ancestors, and put it into the arena.
V2, the second line, adds 3D printing and engineering. A printed shell clips onto the board: plates that slope down as ramps and skirts, or lean up as lips, in PLA or rubbery TPU. Spokes get TPU tyres, legs TPU feet. Magnets are loose diametric cylinders in printed pockets, as in MIT's M-Blocks, so they turn to their neighbour and only ever attract; they sit where robots touch — at a plate's edge, in a foot, near the rim — and their size follows from how hard a robot should hold on: about its own weight, enough to stay up, not so much that it cannot walk away. Wheg spokes are long enough to climb a neighbour's deck (a wheg climbs about 1.5 times its spoke length), and the capacitor holds enough energy for one pulse to lift the robot a good part of its height. [S] on a robot's page gives its printed parts as STL.
V3 builds the motors in as they would really be held: nothing that takes a load is glued. A motor lying on the board sits in a printed cradle, a clamp ring round its can; one standing up sits in a printed sleeve. Each holder has two bosses, and two M2 screws come through holes in the board from its other face and cut their thread in them. A solenoid is screwed on by its steel frame, a coin motor keeps its adhesive pad. The holders lift the motors by their floor, add their weight, take room on the board, and the screw heads on the far face are things a robot can stand or catch on. V3 started from V2's best robots, refitted with holders.
V4 is built from parts that can be ordered: ANYSOLAR solar cells, a Torex voltage detector in place of the long-discontinued 1381, SOT-23 transistors and SMD capacitors that a board house places, and the motors with the values their makers list. The circuit is the classic 1381 solar engine, its latch holding until the capacitor is nearly empty. V4 also learned from the earlier lines: half of its new robots are drawn the way their best robots were built. And it judges heaps more fairly. Tried again from other starts, the earlier leaders scored a fraction of their best run, so a V4 heap is the mean of three starts. And its physics is mended: in V1 to V3 a thin wire leg could be pressed along its length without limit and, when the robot tipped, give back more than it had taken — some vibrating robots were thrown up to 35 cm, and their jump and travel scores owe a part to that. In V4 a wire leg buckles as a real one does, and the spring turning with the robot costs the work it should.
V5 sends twins of strange shapes into one day's sun. V6 gives them honest electrics: solar cells that lose voltage at dawn and in shade, motors with friction, transistors that saturate. V7 gives them honest mechanics. Film strips of V6's best heaps showed robots thrown up to 74 cm, many times their own height, and an energy count found the cause: a gear motor's reaction counted its gearbox's inertia as a flywheel on the robot, so a starting gear spun its robot round, and spokes, walls and wire legs pushed without limit. The heaps of the earlier lines may owe part of their height to the same faults. In V7 no robot has more energy than its motors gave it, and a gear motor's pulse turns its spokes a few degrees, as on the workbench.
V8 makes them small, strange and orderable. Robots that shake are small and light, coin motors ride on slanted bristles like a Hexbug, boards may grow alien outlines, and the motors have the values of parts on sale. Every robot can be ordered: its boards are routed in the browser and come with Gerbers, a BOM and placement files for JLCPCB, built only from parts JLCPCB has.