Your own environments without code: which terrains and challenges do you want?

[mpena2099]

On the next-run branch, coming in the next release, there are two ways to bring your own ground:

• Tune a terrain from the command line. --terrains "steps:rise=0.08:run=0.3,slope:degrees=12" sets the factory's arguments. The archive keeps the string, so replay and validate rebuild the same ground.
• Write a plugin. A plugin is a Python file that registers new terrains, and new tasks for the score, loaded with --plugin my_plugin.py. The archive records the file name and its SHA-256, and every script that opens the archive asks for the same file. The worked example is examples/plugins/progressive_stairs.py: four steps on 28 cm treads, with risers of 4, 8, 12 and 17 cm (a house step).

Two things are still planned. One is drawing the ground as a greyscale PNG (heightmap:yard.png:height=0.05, a MuJoCo height field like rough). The other is storing the whole terrain inside the archive, so a plugin ground opens on another machine without the file.

What I would like to know is which ground is worth supporting next. Which surface would you point at a robot: a doorstep, a gravel path, grass, a loose rug, sand? Which challenge would you add to the score: a gap to cross, a load to carry, a floor it must not damage?

Questions: which one terrain would you ask for before any other, and why that one? If you drew a heightmap yourself, what tool would you use?

Docs: environments, plugins. Roadmap: docs/roadmap.md

[GioiaZheng]

I’d try a short slippery strip on an otherwise flat floor. Unlike slippery, it tests what happens when traction changes mid-stride. Keep the robot and gait fixed, shift the strip between trials, and count how often it stays upright and walks another body length past the far edge.

[mpena2099]

@GioiaZheng Good idea, and it covers something I don't test today. Training varies friction every episode, but always across the whole floor, so a change of grip mid-stride never shows up.

I see it as an evaluation more than a training terrain, the way you described it: keep the robot and its gait, move the strip between trials, count how often it stays upright and walks a body length past the far edge. It fits next to validate.py.

I'll work on it after the next release and post the results here. Thanks!