Environments and presets

Unless a run says otherwise, the robots shown in this documentation were trained on the default set: flat floor, bumps, low steps and rubble (terrain).

Every robot is scored on each environment of the run, and its fitness mixes the average with the worst one (0.7 * mean + 0.3 * min), so the search favours robots that cope with all of them. Pick the environments with --terrains:

python experiments/train.py --terrains rubble            # one environment only
python experiments/train.py --terrains flat,soft,steep   # your own list
python experiments/train.py --terrains terrain           # a ready set (below)

Fewer environments make a specialist (the best robot for sand, or for a staircase); more make a generalist. The choice is saved in the archive, so replay, validate and export use the same ones.

Name	What the robot faces
flat	flat floor
rough	smoothed random bumps up to 4 cm, flat near the spawn
rubble	40 scattered rocks 3 to 7 cm high, taller than many wheels can climb
steps	six low steps, 3 cm high and 50 cm deep, the first at x = 0.6 m (see below)
ledges	staircase of four 6 cm steps, too tall to roll up
step	one 3 cm step, 0.50 m ahead of the spawn, 1 m deep
slope	8 degree ramp up from x = 0.5 m
steep	the same ramp at 15 degrees
downhill	spawn on a plateau, then a 10 degree slope down to the floor
slippery	floor with friction 0.2, like ice
soft	yielding floor like sand: feet sink 1 to 2 cm and wheels feel rolling drag

How hard steps is. It is not a house staircase: each step is 3 cm high (a house step is 17 to 18 cm, taller than most of these robots) and 50 cm deep. In an 8 s episode a good robot covers 1.5 to 2.3 m, so it meets the first two to four steps, at x = 0.6, 1.1, 1.6 and 2.1 m. A wheel climbs an edge up to about its own radius, and wheels here are 30 to 60 mm, so a low robot on wheels rolls up a 3 cm step; a low body without wheels can push its torso over the edge like a sled. Robots trained only on flat ground still mostly stall there. ledges doubles the height to 6 cm, too tall to roll up.

Ready sets:

Set	Environments
flat	flat
terrain	flat, rough, steps, rubble
all	all of the above except step

rough and rubble are random. Training draws new bumps and rocks for every body, so a family of robots cannot memorise one layout; replay and validate use one fixed layout, so a robot always meets the same ground there.

Make your own environment

An environment is a function in darwinbot/terrain.py that returns a Terrain: a name, the MuJoCo XML of the world (boxes, cylinders, meshes, anything MuJoCo loads), the ground friction and the ground height at the spawn. Add it to TERRAINS and train on it:

python experiments/train.py --terrains tunnel          # only the new one
python experiments/train.py --terrains flat,tunnel     # mixed with others

Evolution then keeps the bodies and gaits that do best there. A few things decide whether the result is what you expect:

• Episodes are short. Each test lasts 8 s and good robots cover about 1.5 m, so the challenge has to sit within the first metre or two after the spawn (x = 0, robots head towards +x).
• The score stays the same. It rewards distance along x, turning when told to, low energy and not falling. A tunnel rewards going through it forwards; reaching a goal or solving a maze would need a new task in fitness.py, and the robots have no sensors to find their way.
• A fixed layout gets memorised. With every object always in the same place, evolution learns that one layout. Draw the layout from a seed, as rough and rubble do, so every body meets a new one.
• The map is still measured on flat ground. The archive's axes (contact time, height, legs, cost of transport) come from the flat floor; a new environment changes the fitness, not how robots are sorted.
• Contacts cost time. rubble has 40 rocks; hundreds of objects can make training several times slower.
• The world can touch the robot at the spawn. The spawn check only looks for parts of the robot touching each other, so a ceiling lower than a body pins it in place, which in practice filters tall bodies out.

Tasks and presets

The ready-made tasks and the --preset option are described in Fitness and tasks.