First run

Train a small archive, list the robot types it found and watch one move.

python experiments/train.py
python experiments/replay.py --list
python experiments/replay.py --type 8servos-mid-trot

All commands

# Evolve on the flat floor with the default settings
python experiments/train.py

# Flat, rough, steps and rubble, with more batches and workers
python experiments/train.py --terrains terrain --batches 60 --workers 8

# Learn to turn in place instead of walking forward
python experiments/train.py --task turn

# Only some environments (see Environments and presets)
python experiments/train.py --terrains rubble,soft

# Climb a 3 cm step, or carry 0.3 kg on the torso
python experiments/train.py --terrains step
python experiments/train.py --payload 0.3

# Resume a previous run (task and payload come from the archive)
python experiments/train.py --resume checkpoints/archive.pkl --batches 20

# Draw parents by novelty (default) or uniformly at random
python experiments/train.py --parent-selection uniform

# Replay the best elite in the MuJoCo viewer (flat terrain)
python experiments/replay.py

# List the robot types found so far, then open one of them
python experiments/replay.py --list
python experiments/replay.py --type 8servos-mid-trot --terrain rough

# Record a video instead of opening a window (needs the video extra)
python experiments/replay.py --record out.mp4 --seconds 8

# Train on randomised models (default 2 per terrain on cpu, 0 on gpu)
python experiments/train.py --variants 4

# Check the elite on 50 randomised models per terrain (PASS if mean loss < 30%)
python experiments/validate.py --n 50

# Export a robot to build: parts, bill of materials, assembly, control file
python experiments/export.py --type 8servos-mid-trot

# Plot the behaviour scatter and the progress curves
python experiments/plot_archive.py
python experiments/plot_archive.py --show

# Family tree: ancestors of one robot, or ancestor variety against fitness
python experiments/lineage.py --type 8servos-mid-trot
python experiments/lineage.py

Training writes checkpoints/archive.pkl after every batch. replay.py reads that file by default (--archive changes it). plot_archive.py saves archive_heatmap.png (the behaviour scatter plot: one dot per elite at its duty and height, coloured by fitness, sized by servo count) and archive_history.png to checkpoints/ (--out changes the directory).

--variants N scores every candidate gait on N randomised copies of each terrain (friction, mass, servo torque and stiffness, joint play, 0 to 2 control periods of command delay) and takes the mean, so gaits that depend on exact simulator details lose. It costs about N times more per evaluation, is CPU only, and is stored in the archive. validate.py runs the elite on --n fresh variants (--seed, --workers) and prints nominal, mean and worst distance, fall rate and relative loss per terrain.

Viewing the different robot types

The archive keeps one robot per cell of the behaviour map: the best one found with that duty, height, number of legs and cost of transport (four descriptors). A new robot only replaces the one in its own cell, and only if it scores higher, so slow crawlers survive next to fast rollers.

A type is a short name derived from the cell, <servos>servos-<height>-<gait> (for example 8servos-mid-trot):

Part	Values	Meaning
servos	1 to 16	number of servos, the cost of the robot (shown in the name, not a map axis)
height	low, mid, high	mean torso height over body reach: below 0.4, below 0.8, above
gait	crawl, walk, trot, belly, roll	duty, the share of time the tips touch the ground: above 0.8, 0.5 to 0.8, 0.15 to 0.5, below 0.15 (sliding on the body); roll when wheels carry more than half of the contact time

Several cells can share a name; the later ones get the cell index appended (12servos-low-walk-133).

Step 1, see which types the run has found, best first:

$ python experiments/replay.py --list
type                       fitness  duty  height  servos    cot
8servos-mid-trot              0.11  0.34    0.45       8    6.2
12servos-low-trot            -0.03  0.41    0.28      12   21.4
12servos-low-walk-133        -0.03  0.52    0.33      12   18.7
...

Step 2, open one in the 3D viewer (or add --record name.mp4 for a video):

python experiments/replay.py --type 8servos-mid-trot

Without --type, replay.py shows the best robot overall.

Viewer controls

replay.py without --record opens the interactive MuJoCo viewer: left drag rotates the camera, right drag pans, scroll zooms, double click selects a body and Ctrl + double click makes the camera follow it, space pauses, Tab hides the side panels. --speed 0.25 plays in slow motion (the window has no speed control of its own). The arrow keys steer the robot while it runs (left and right turn, up or down go straight), and --command -1 or 1 starts it turning, also for --record. The simulation does not reset, so the robot eventually leaves the terrain.