Environment Reference

• CPU — make_env(env_id) (from myosuite import make_env). Playback and SB3.

• mjlab — the same env_id for GPU training (scripts/train_mjlab.py), or from Python with make_env(env_id, backend="mjlab", num_envs=...).

List every CPU ID on your install:

python -c "import myosuite; print('\n'.join(myosuite.myosuite_env_suite))"

Tables below are the common CPU IDs. Muscle-condition variants (myoSarc…, myoFati…, hand myoReaf…) are registered for the myo* tasks; see Naming Conventions.

GPU (mjlab) coverage

The mjlab backend registers a twin under the same env_id for these CPU families (myoSarc…/myoFati…/myoReaf… variants exist where the CPU env has them):

Family	mjlab twins
Pose	myoElbowPose1D6M{Fixed,Random} (+ Exo), myoFingerPose{Fixed,Random}, myoHandPose{0-9}Fixed, myoHandPose{Fixed,Random}, myoTorsoPoseFixed, myoTorsoExoPoseFixed, motorFingerPose{Fixed,Random}
Reach	myoArmReach, myoFingerReach, myoHandReach, motorFingerReach (each {Fixed,Random})
Leg	myoLegWalk, myoLegDirectional{Forward,Backward,Random}, myoLegStandRandom, myoLeg{Rough,Hilly,Stair}TerrainWalk
Challenge	myoChallengeChaseTagFBP2, myoChallengeTableTennisP{0,1,2}

Variants: myoSarc… exist for all twins above except myoLegDirectional* (no CPU variant) and the Challenge twins; myoFati… likewise, and myoReaf… for the hand pose and reach twins.

MuscleMimic (myoMimicFullbody-v0, myoMimicBimanual-v0 and their SAR variants) has its own mjlab implementation. It is not built from the CPU registration and is registered once a motion clip is given (MYOSUITE_MIMIC_CLIP or register_mimic_mjlab_tasks_with_clip; see tutorials 5.x), and it takes no muscle-command features.

Not on mjlab yet: the other Challenge tasks, myoChallengeChaseTagFBVs (two-agent), myoElbowPoseTask*, myoFullBodyDirectional and the hand manipulation families (myoHandKeyTurn, ObjHold, PenTwirl, Reorient*) including their variants. The cross-backend contract lists which envs and backends take which features.

Differences from the CPU env that matter when moving policies between backends:

• Torso exosuit (myoTorsoExoPoseFixed and variants) — experimental: the two free exosuit bodies are 6-DoF joint chains on mjlab (mjlab allows one freejoint per entity); their qpos/qvel are converted back to the CPU layout, so the observation is the same 296-d vector (matching CPU to ~1e-3 over short rollouts, see test_torso_exo_observation_matches_cpu). Transfer of trained policies between the backends has not been tested yet.

• Terrain walks: the terrain is baked into the model with a fixed seed instead of being resampled at each reset (myoLegRoughTerrainWalk uses one fixed sample).

• Leg locomotion twins (walk, directional, terrain) are step-parity tested like the other twins, with looser tolerances where contacts dominate (see Backend parity). Height-field contacts differ most, because MuJoCo Warp creates at most one capsule-hfield contact.

Naming Conventions

Base environment IDs follow the pattern:

myo<Model><Task>[Difficulty]-v<N>

Muscle-condition variants are registered for the base myo* CPU environments. Each variant is the base env registered together with the matching wrapper (SarcopeniaWrapper, FatigueWrapper, ReafferentationWrapper), so the same features work on any env that runs muscle stages; see Neuroscience Quick Start for examples and docs/wiki/cross-backend-contract.md for how features are registered and which envs and backends support them:

Prefix pattern	Condition	Example
myoSarc<…>	Sarcopenia (50 % peak force)	myoSarcElbowPose1D6MRandom-v0
myoFati<…>	Cumulative neuromuscular fatigue	myoFatiHandPoseRandom-v0
myoReaf<…>	Tendon transfer / reafferentation (all hand envs)	myoReafHandPoseFixed-v0

The suffix Fixed indicates a fixed (non-random) target; Random indicates a randomly sampled target each episode.

myoFinger (4 DoF, 5–6 muscles)

Environment ID	Task	Diff.	Sarc	Fati
myoFingerReachFixed-v0	Fingertip reach	Easy	✓	✓
myoFingerReachRandom-v0	Fingertip reach	Hard	✓	✓
myoFingerPoseFixed-v0	Joint pose	Easy	✓	✓
myoFingerPoseRandom-v0	Joint pose	Hard	✓	✓

Each base environment also exposes myoSarc… and myoFati… variants (12 total IDs for this model).

myoElbow (1 DoF, 6 muscles)

Environment ID	Task	Diff.	Sarc	Fati
myoElbowPose1D6MFixed-v0	Joint pose	Easy	✓	✓
myoElbowPose1D6MRandom-v0	Joint pose	Hard	✓	✓
myoElbowPose1D6MExoFixed-v0	Pose + elbow exoskeleton	Easy	✓	✓
myoElbowPose1D6MExoRandom-v0	Pose + elbow exoskeleton	Hard	✓	✓

Each row additionally has myoSarc… and myoFati… variants (12 total IDs).

Experimental, defined by a TaskConfig instead of an env class (tutorial 4.3): myoElbowPoseTaskFixed-v0 and myoElbowPoseTaskRandom-v0, with myoSarc… and myoFati… variants.

myoHand (23 DoF, 39 muscles)

Environment ID	Task	Diff.	Sarc	Fati	Reaf
myoHandPoseFixed-v0	23-DoF joint pose	Easy	✓	✓	✓
myoHandPoseRandom-v0	23-DoF joint pose	Hard	✓	✓	✓
myoHandReachFixed-v0	Fingertip spatial reach	Easy	✓	✓	✓
myoHandReachRandom-v0	Fingertip spatial reach	Hard	✓	✓	✓
myoHandKeyTurnFixed-v0	Key rotation (thumb + index)	Easy	✓	✓	✓
myoHandKeyTurnRandom-v0	Key rotation, random init	Hard	✓	✓	✓
myoHandObjHoldFixed-v0	Object repositioning (no drop)	Easy	✓	✓	✓
myoHandObjHoldRandom-v0	Random object, random target	Hard	✓	✓	✓
myoHandPenTwirlFixed-v0	Pen twirl to fixed orientation	Easy	✓	✓	✓
myoHandPenTwirlRandom-v0	Pen twirl to random orientation	Hard	✓	✓	✓
myoHandReorient8-v0	Reorient 1 of 8 objects	Easy	✓	✓	✓
myoHandReorient100-v0	Reorient 1 of 100 objects	Medium	✓	✓	✓
myoHandReorientID-v0	Reorient 1 of 1000 (in-domain)	Hard	✓	✓	✓
myoHandReorientOOD-v0	Reorient 1 of 1000 (out-of-domain)	Hardest	✓	✓	✓

With all three variants each row generates myoSarc…, myoFati…, and myoReaf… IDs. The full hand environment count is 14 × 4 = 56 IDs.

MyoChallenge hand tasks (several also register myoSarc…/myoFati… variants, 40 variant IDs in total):

Environment ID	Task
myoChallengeDieReorientDemo-v0	Die reorientation demo
myoChallengeDieReorientP1-v0	Die reorientation (limited goal range)
myoChallengeDieReorientP2-v0	Die reorientation (full range + friction/size variation)
myoChallengeBaodingP1-v1	Baoding balls — swap positions
myoChallengeBaodingP2-v1	Baoding balls — full rotation + size/friction variation
myoChallengeRelocateP1-v0	Grasp & place object (phase 1)
myoChallengeRelocateP2-v0	Grasp & place (phase 2, harder)
myoChallengeRelocateP2eval-v0	Relocate phase 2 — evaluation split
myoChallengeTableTennisP0-v0	Table tennis swing (warm-up)
myoChallengeTableTennisP1-v0	Table tennis (phase 1)
myoChallengeTableTennisP2-v0	Table tennis (phase 2, full task)
myoChallengeBimanual-v0	Bimanual object manipulation

myoArm (hand-free variant used here: 18 DoF, 50 muscles; full arm: 27 DoF, 63 muscles)

Environment ID	Task	Diff.	Sarc	Fati
myoArmReachFixed-v0	Index fingertip reach (fixed)	Easy	✓	✓
myoArmReachRandom-v0	Index fingertip reach (random)	Hard	✓	✓

The arm model used here is the hand-free variant (extrinsic + intrinsic hand muscles removed) to isolate reaching without manipulation.

myoLeg (14 joints plus the 6-DoF free root = 20 DoF, 80 muscles)

Environment ID	Task	Diff.	Sarc	Fati
myoLegStandRandom-v0	Static balance — random init	Easy	✓	✓
myoLegWalk-v0	Flat-ground forward walking	Medium	✓	✓
myoLegRoughTerrainWalk-v0	Walking on rough terrain	Hard	✓	✓
myoLegHillyTerrainWalk-v0	Walking on hilly terrain	Hard	✓	✓
myoLegStairTerrainWalk-v0	Stair climbing	Hardest	✓	✓
myoLegDirectionalForward-v0	Directional walking — forward	—	—	—
myoLegDirectionalBackward-v0	Directional walking — backward	—	—	—
myoLegDirectionalRandom-v0	Directional walking — random direction	—	—	—

MyoChallenge leg / whole-body tasks:

Environment ID	Task
myoChallengeChaseTagP1-v0	Chase-tag locomotion (phase 1)
myoChallengeChaseTagP2-v0	Chase-tag (phase 2, two-agent)
myoChallengeChaseTagP2eval-v0	Chase-tag phase 2 — evaluation split
myoChallengeOslRunFixed-v0	OSL prosthetic running (fixed terrain)
myoChallengeOslRunRandom-v0	OSL prosthetic running (random terrain)
myoChallengeSoccerP1-v0	Soccer ball kicking (phase 1)
myoChallengeSoccerP2-v0	Soccer ball kicking (phase 2)

myoTorso (18 joints, 216 muscles)

Environment ID	Task	Diff.	Sarc	Fati
myoTorsoPoseFixed-v0	Lumbar spine pose	Fixed	✓	✓
myoTorsoExoPoseFixed-v0	Lumbar spine pose + exoskeleton	Fixed	✓	✓

Full body and MuscleMimic

Environment ID	Task	mjlab
myoMimicFullbody-v0 (also myoMuscleMimicFullbody-v0)	MuscleMimic full-body motion imitation	yes
myoMimicBimanual-v0 (also myoMuscleMimicBimanual-v0)	MuscleMimic bimanual arm imitation	yes
myoFullBodyDirectional-v0	Full-body directional locomotion	no
myoChallengeChaseTagFBP2-v0	Chase-tag, full body, scripted opponent	yes
myoChallengeChaseTagFBVs-v0	Chase-tag, full body, two agents (1v1)	no

The MuscleMimic and mjlab columns are explained under “GPU (mjlab) coverage” above.

mjlab (GPU)

Install pip install -e ".[mjlab]" (see Installation for matching the torch build to your driver’s CUDA version) and train with the CPU env_id:

python scripts/train_mjlab.py myoElbowPose1D6MRandom-v0 --env.scene.num-envs 1024

List the registered mjlab IDs with mjlab.tasks.registry.list_tasks().

MJX (experimental)

A JAX path exists (pip install -e ".[mjx]", from myosuite.envs.myo.backends.mjx import make). Do not start new work on it; use mjlab for GPU training. The MJX envs are not observation- or reward-compatible with the CPU/mjlab envs of the same task, so MJX-trained policies do not transfer to them (see the limitations in myosuite/envs/myo/backends/mjx/README.md).

Supported envs:

Family	MJX ids (make("<id>"))
Pose	MjxElbowPoseFixed-v0, MjxElbowPoseRandom-v0, MjxFingerPoseFixed-v0, MjxFingerPoseRandom-v0, MjxHandPoseRandom-v0
Reach	MjxHandReachFixed-v0, MjxHandReachRandom-v0, MjxFingerReachRandom-v0
Walk	MjxLegWalk-v0 (flat ground, mirrors myoLegWalk-v0)
Mimic	MjxMimicBimanual-v0, MjxMimicFullbody-v0 (aliases MjxMuscleMimicBimanual-v0, MjxMuscleMimicFullbody-v0)

MJX ids use the Mjx prefix and are not the myo… ids of the CPU and mjlab backends; there is no MJX version of the torso, arm, leg terrain/stand/directional envs or of the myoSarc…/myoFati…/myoReaf… variants. The list is defined in myosuite/envs/myo/backends/mjx/__init__.py.