Defining an Environment

In MyoAssist, an environment is a human musculoskeletal (MSK) model, an assistive device, and a terrain, composed into one MuJoCo model. The same definition drives both pipelines: reflex Controller Optimization (CO) and Reinforcement Learning (RL).

The environment spec

An environment has three fields. Each field is a raw registry key:

{ "msk": "myolegs22", "device": "Humotech_L1", "terrain": { "terrain": "slope", "deg": 8 } }

Field	What it is	Examples
msk	human MSK model	myolegs22, myolegs26
device	assistive device	DephyExoBoot_L1, Humotech_L1, OpenSourceLeg_A_L1, Tutorial_L1
terrain	the ground (optional)	omit for flat; an inline config; or a terrains JSON path

Run python -m assist_sim list to see every MSK and device that is installed, and which pairs are compatible.

MSK models

Key	Muscles	Description
myolegs22	22	2D (sagittal-plane)
myolegs26	26	3D
myolegs	80	3D. The key of the 80-muscle model is myolegs, not myolegs80.
myofullbody	416	3D, full body with arms and torso muscles

The muscle count and the 2D or 3D control mode come from msk. You do not set them separately.

For a 3D key, the CO pipeline replaces the free root with named sagittal DOFs, because the reflex controller reads the pelvis state from them. This is automatic. See Root frame.

Discovering and validating

python -m assist_sim list prints the authoritative installed set. python -m assist_sim validate <msk> <device> checks one pair.

python -m assist_sim list                       # every installed MSK / device + compatibility
python -m assist_sim validate myolegs22 Humotech_L1

EnvSpec.validate() does the same check in code. It raises a ValueError that lists the valid options when a key is unknown or the MSK and device pair is incompatible.

Note:

Environment validation must use the correct raw registry keys or the returned result will be inaccurate.

Example:

python -m assist_sim validate myolegs22 Dephy_L1
INVALID: myolegs22 x Dephy_L1

python -m assist_sim validate myolegs22 DephyExoboot_L1
INVALID: myolegs22 x DephyExoboot_L1

python -m assist_sim validate myolegs22 DephyExoBoot_L1
OK: myolegs22 x DephyExoBoot_L1
   human:  myolegs22 (composed MjSpec, 38 bodies)
   config: ...\Lib\site-packages\assist_sim\models\DephyExoBoot\L1config.yaml

Terrain

Leave terrain unset (or null) for a flat, effectively-infinite ground plane. Otherwise give one of the following.

A uniform surface (one plane or one heightfield):

terrain	Result
{ "terrain": "flat" }	a flat plane
{ "terrain": "slope", "deg": 8 }	a constant 8° incline (a tilted plane)
{ "terrain": "random", "amplitude": 0.06 }	a rough heightfield, up to 6 cm relief
{ "terrain": "sinusoidal", "amplitude": 0.05, "period": 1.0 }	rolling waves

Those are the common cases. The full field set, including resolution, extent, safe_zone_radius and seed, is in Uniform Terrains. Read resolution and extent together: they set the heightfield cell size, which is what the roughness actually feels like underfoot.

A tiled grid: a terrain config with a grid and per-cell tiles. Tile types are flat, slope, stairs, pyramid_stairs, rough, boulders, stepping_stones, discrete_obstacles, and gap. You can fill empty cells with randomization. Give the config inline or as a path to a JSON file:

"terrain": {
  "grid": { "rows": 1, "cols": 3, "tile_size": [4.0, 4.0] },
  "border": { "width": 0.5 },
  "tiles": [
    { "row": 0, "col": 0, "type": "flat",   "params": { "height": 0.0 } },
    { "row": 0, "col": 1, "type": "slope",  "params": { "angle_deg": 8.0, "axis": "x" } },
    { "row": 0, "col": 2, "type": "stairs", "params": { "n_steps": 5, "step_height": 0.1, "axis": "x" } }
  ]
}

The terrain sets the course grade. A slope terrain is the incline, and the evaluation camera, cost, and readouts derive the angle from it. There is no separate slope flag.

Reflex CO is for steady-state locomotion. A constant terrain is manageable. The CO pipeline and reflex controller do not optimize over highly variable terrain (rough, stairs, mixed tiles). Use variable terrain with RL.

Using an environment spec

Controller Optimization (reflex)

Run these from the repository root — the --env-spec path is repo-root-relative. python -m ctrl_optim.optim.train … takes ad-hoc flags; to run a named config instead, use run_optim.py (see Running Optimizations).

# raw flags
python -m ctrl_optim.optim.train --msk myolegs22 --device Humotech_L1 \
    --terrain '{"terrain":"slope","deg":8}' --sim_time 20 -eff --ExoOn 1

# or a shared env-spec file
python -m ctrl_optim.optim.train --env-spec docs/examples/env_exo_slope.json --sim_time 20 -eff --ExoOn 1

Reinforcement Learning

Set the same three fields on env_params in your training-config JSON:

"env_params": { "msk_key": "myolegs22", "device_key": "Humotech_L1", "terrain": null }

Programmatically

from myoassist_utils.env_spec import EnvSpec

spec = EnvSpec.load("docs/examples/env_exo_slope.json")
# or: EnvSpec(msk="myolegs22", device="Humotech_L1", terrain={"terrain": "slope", "deg": 8})

spec.validate()          # checks keys against the registry; raises with valid options on a bad key
xml = spec.compose()     # returns a loadable MuJoCo MJCF string
spec.compose(export_path="my_env.xml")   # also writes a standalone, loadable file

Ready-to-use examples

The docs/examples/ directory in the myoassist repo has runnable env-specs:

File	Environment
env_exo_flat.json	myolegs22 + Humotech_L1, flat
env_exo_slope.json	myolegs22 + Humotech_L1, 8° slope
env_prosthesis_rough.json	myolegs22 + OpenSourceLeg_A_L1, rough heightfield
env_tiled_course.json	myolegs22 + Humotech_L1, a flat, slope, and stairs tiled course
env_tiled_random.json	myolegs22 + OpenExo_L1, a randomized 3×3 tiled grid