Fitness and tasks

Tasks

--task picks what the score rewards; the body and controller genomes are the same for every task, and so are the behaviour descriptors.

• steer (default): walk straight and turn when told. A steering command from -1 (left) to +1 (right) scales each joint's amplitudes, shifts its offset and changes each wheel's speed through two genes per slot, the way animals turn with one gait and one side stronger. Every evaluation runs three episodes, straight, left and right: score = 0.5 * straight + 0.25 * (left + right), where straight is the forward score and a turn scores 0.2 * yaw towards the side asked for (yaw in radians, counted up to 90 degrees), never more than the distance the torso travelled in that episode, minus the same energy and fall terms. The robot has to walk a curve: spinning on the spot earns nothing. Three episodes make an evaluation three times as slow. On the robot the command can come from a remote control, and later from sensors.
• forward: walk straight only, the score above.
• turn: rotate in place. score = |yaw| - 0.5 * hypot(dx, dy) - 0.05 * energy / (m * g) - 0.5 * fell, where yaw is the net heading change of the torso in radians, either way.

--payload KG adds mass to the torso payload box for the whole run. The task and payload are stored in the archive, so replay.py, validate.py and export.py use the same conditions the robot was trained under (replay.py --list prints them in a header line when they are not the default).

Ready-made tasks run with --preset, one after the other, each in its own folder (checkpoints/<preset>), so they never overwrite the main archive:

Preset	Task	Terrain	Payload
steer	walk and turn when told	the terrain set	none
forward	walk forward	the terrain set	none
turn	turn in place	flat	none
step	walk and turn when told	one 3 cm step	none
payload	walk and turn when told	the terrain set	0.3 kg

# The three extra tasks, 5 batches each. step and payload score three episodes
# per evaluation, so each takes about 10 h on 4 cores; turn is much quicker
python experiments/train.py --preset turn,step,payload --batches 5 --workers 3
# The same command again continues each preset from where it stopped
python experiments/replay.py --archive checkpoints/turn/archive.pkl --list

Ctrl+C saves the preset being trained and skips the rest.

Fitness

Per episode, from darwinbot/fitness.py:

score = dx - 0.5 * |dy| - 0.2 * |yaw| - 0.05 * energy / (m * g) - 0.5 * fell

dx is forward displacement, dy lateral displacement, yaw the net heading change in radians (charged at the rate a turn earns it), energy the mechanical work of all actuators, m the robot mass and g = 9.81. Dividing energy by weight turns it into metres and keeps heavy and light bodies comparable. A fall ends the episode but keeps the distance covered, so the fixed penalty only breaks ties in favour of staying upright. A physics blow-up returns UNSTABLE_SCORE = -5.0.

Multi-terrain runs use a blend of mean and worst score:

fitness = 0.7 * mean(scores) + 0.3 * min(scores)