Models and Tasks

Models and tasks are the most important aspects of Myosuite. To provide further,

• Models

  • myoFinger

  • myoElbow

  • myoHand

  • myoLeg

  • myoArm

• Tasks

• Task and Variantions

Models

MyoSuite consists of five models: myoFinger, myoElbow, myoHand, myoArm and myoLeg. Using these models, we design a rich collection of tasks ranging across simple reaching movements, contact-rich movements involving object-manipulation such as pen-twirling and baoding balls, as well as locomotion behaviors.

myoFinger

Simplified and intuitive model (based on Xu et al. (2012)[https://ieeexplore.ieee.org/document/6290710]) of a 4 Degree of Freedom (DoF) finger which is actuated through a series of 5 simplified antagonistic muscle-tendon units. We also provide its robotic counterpart with simple torque actuators to facilitate the comparative investigation.

Details of this model can be found at myoFinger

Acronym	Muscle
EXTN	Central Extensor
adabR	Abduction Tendon(Right)
adabL	Abduction Tendon(left)
mflx	Pip Flexor
dflx	DIP Flexor

myoElbow

A model of 2 DoF human elbow joint – based on OpenSim’s default testing arm model (Delp et al. (2007); Seth et al. (2018)) – and actuated using multiple agonist/antagonist pairs (3 flexors and 3 extensors).

Details of this model can be found at myoElbow

Acronym	Muscle
TRIlong	Triceps long
TRIlat	Triceps lateral
TRImed	Triceps medial
BIClong	Biceps long
BICshort	Biceps short
BRA	Brachioradialis

myoHand

The dexterous human hand requires coordination of multiple highly redundant muscles, which have complementary and antagonistic effects on various joints. This musculoskeletal model is comprised of 29 bones, 23 joints, and 39 muscles-tendon units.

Details of this model can be found at myoHand

Acronym	Muscle
ECRL	Extensor Carpis Radialis Longus
ECRB	Extensor Carpis Radialis Brevis
ECU	Extensor Carpi Ulnaris
FCR	Flexor Carpi Radialis
FCU	Flexor Carpi Ulnaris
PL	Palmaris longus
PT	Pronator teres
PQ	Pronator
EIP	Extensor Indicis Proprius
EPL	Extensor Pollicis Longus
EPB	Extensor Pollicis Brevis
FPL	Flexor Pollicis Longus
APL	Abductor Pollicis Longus
OP	Opponens Pollicis
FDS	Flexor Digitorum Superficialis (2- index, 3- middle, 4- ring, 5- little)
FDP	Flexor Digitorum Profundus (2- index, 3- middle, 4- ring, 5- little)
EDC	Extensor Digitorum Communis (2- index, 3- middle, 4- ring, 5- little)
EDM	Extensor Digiti Minimi
RI	Radial Interosseous (2- index, 3- middle, 4- ring, 5- little)
LU-RB	Lumbrical (2- index, 3- middle, 4- ring, 5- little)
UI-UB	Palmar or Ulnar Interosseous (2- index, 3- middle, 4- ring, 5- little)

myoLeg

The 3D dexterous human leg requires coordination of multiple highly redundant muscles, which have complementary and antagonistic effects on various joints.

This musculoskeletal model is comprised of 10 joints, 20 DoFs, and 80 muscles-tendon units. This musculoskeletal model takes Rajagopal’s full body gait model [https://github.com/opensim-org/opensim-models/tree/master/Models/RajagopalModel] as close reference.

Joints and muscle details can be found in Rajagopal’s paper [https://ieeexplore.ieee.org/document/7505900]

The model implementation details can be found on our github repo at myoLeg

Acronym	Muscle
addbrev_r	Adductor brevis
addlong_r	Adductor longus
addmagDist_r	Adductor magnus (distal)
addmagIsch_r	Adductor magnus (ischial)
addmagMid_r	Adductor magnus (middle)
addmagProx_r	Adductor magnus (proximal)
bflh_r	Biceps femoris long head
bfsh_r	Biceps femoris short head
edl_r	Extensor digitorum longusc
ehl_r	Extensor hallucis longusc
fdl_r	Flexor digitorum longus
fhl_r	Flexor hallucis longus
gaslat_r	Gastrocnemius lateral head
gasmed_r	Gastrocnemius medial head
glmax1_r	Gluteus maximus (superior)
glmax2_r	Gluteus maximus (middle)
glmax3_r	Gluteus maximus (inferior)
glmed1_r	Gluteus medius (anterior)
glmed2_r	Gluteus medius (middle)
glmed3_r	Gluteus medius (posterior)
glmin1_r	Gluteus minimus (anterior)
glmin2_r	Gluteus minimus (middle)
glmin3_r	Gluteus minimus (posterior)
grac_r	Gracilis
iliacus_r	Iliacus
perbrev_r	Peroneus brevis
perlong_r	Peroneus longus
piri_r	Piriformis
psoas_r	Psoas
recfem_r	Rectus femoris
sart_r	Sartorius
semimem_r	Semimembranosus
semiten_r	Semitendinosus
soleus_r	Soleus
tfl_r	Tensor fascia latae
tibant_r	Tibialis anterior
tibpost_r	Tibialis posterior
vasint_r	Vastus intermedius
vaslat_r	Vastus lateralis
vasmed_r	Vastus medialis

myoArm

The 3D dexterous human arm reaching and manipulation.

This musculoskeletal model is comprised of 27 Degree of Freedom, and 63 muscles-tendon units.

Model specification can be found on our github repo of myoArm

Tasks

Finger Joint Pose

Objective:

• Strike a joint pose

Variants:

• Easy: Move to a fixed specified joint pose myoFingerPoseFixed-v0

• Hard: Move to randomly selected joint poses myoFingerPoseRandom-v0

• More: Random resets, no resets, dense rewards, sparse rewards

Finger Tip Reach

Objective:

• Reach using finger tips

Variants:

• Easy: Reach to a fixed location myoFingerReachFixed-v0

• Hard: Reach to random locations myoFingerReachRandom-v0

• More: Random resets, no resets, dense rewards, sparse rewards

Elbow Joint Pose

myoElbow model with 6 muscles (3 flexors and 3 extensors) was simplified to have only elbow rotations. Although it is not a physiologically accurate model it can be a very simple model for troubleshooting initial control schemes.

Objective:

Move elbow to a specified pose.

Variants:

• Easy: Move to random joint poses using 1 muscle ElbowPose1D1MRandom-v0

• Hard: Move to random joint poses using 6 muscles myoElbowPose1D6MRandom-v0

• More: Random resets, no resets, dense rewards, sparse rewards

Hand Joints Pose

Drive the myoHand model to achieve a joint poses. In addition to making co-ordinated movements, avoiding self collisions poses additional challenges in solving this task.

Objective:

Strike a hand pose

Variants:

• Easy: Move to a fixed joint pose myoHandPoseFixed-v0

• Hard: Move to a randomly selected joint pose myoHandPoseRandom-v0

• More: Random resets, no resets, dense rewards, sparse rewards

Hand Tips Reach

Make reaching movements using the myoHand model to have finger tips achieve a specific spatial position. In addition to making co-ordinated movements, avoiding self collisions poses additional challenges in solving this task.

Objective:

Reach using finger tips

Variants:

• Easy: Reach fixed positions using finger tips myoHandReachFixed-v0

• Hard: Reach random positions using finger tips myoHandReachRandom-v0

• More: Random resets, no resets, dense rewards, sparse rewards

Hand Key Turn

Use the the myoHand model to rotate a key. The complexity of this task is due to the use of only index and thumb fingers to solve the task with intermittent contact with the key.

Objective:

Coordinate finger movements to rotate a key

Variants:

• Easy: Achieve half rotation of the key myoHandKeyTurnFixed-v0

• Hard: Achieve full rotation + random initial configuration myoHandKeyTurnRandom-v0

• More: Random resets, no resets, dense rewards, sparse rewards

Hand Object Hold

A myoHand model moves an object in the hand to a given orientation without dropping it. The complexity of this task is due to the intermittent contacts between the object and multiple fingers needing co-ordination to stabilize the object.

Objective:

Reposition an object to reach a given target without dropping it.

Variants:

• Easy: Reposition to a fixed position myoHandObjHoldFixed-v0

• Hard: Reposition a random object to random positions myoHandObjHoldRandom-v0

• More: sparse rewards, dense rewards, random resets, reset free

Hand Pen Twirl

A myoHand model rotates a pen in the hand to a given orientation without dropping it. The complexity of this task is due to the intermittent contacts between the object and multiple fingers while trying to stabilize the object.

Objective:

Rotate the object to reach a given orientation (indicated by the green object in the scene) without dropping it.

Variants:

• Easy: Rotate to fixed orientation myoHandPenTwirlFixed-v0

• Hard: Rotate of random orientation myoHandPenTwirlRandom-v0

• More: sparse rewards, dense rewards, random resets, reset free

Hand Multiobject Reorientation

A myoHand model reorients a set of parameterized geometries to a given target orientation without dropping it. As with before, the complexity of this task is due to the intermittent contacts between the object and multiple fingers while trying to stabilize the object, as well as the multiobject nature of the task.

Objective:

Rotate one randomly chosen object in the set to reach a given orientation without dropping it.

Variants:

• Easiest: Eight-object reorientation myoHandReorient8-v0

• Easy: 100-object reorientation myoHandReorient100-v0

• Hard: 1000-object orientation (in-domain parameterized geoms) myoHandReorientID-v0

• Hardest: 1000-object orientation (out-of-domain parameterized geoms) myoHandReorientOOD-v0

Hand Die Rotation (MyoChallenge 2022)

A myoHand model involved for reconfiguring a die to match desired goal orientations. This task require delicate coordination of various muscles to manipulate the die without dropping it.

Objective:

Rotate the object to reach a given orientation (indicated by the green reference object in the scene) without dropping it.

Action Space:

The whole set of muscle [0, 1]

Observation Space:

All joints angles [-\(\pi\), \(\pi\)]

Variants:

• Easy: Goal position and orientation limited myoChallengeDieReorientP1-v0

• Hard: Large range of goal position and rotations. Variability in object size and frictions. myoChallengeDieReorientP2-v0

• More: Sparse rewards, 3 different dense reward options to choose from

Hand Baoding Balls (MyoChallenge 2022)

A myoHand model involved for simultaneous rotation of two free-floating spheres over the palm. This task requires both dexterity and coordination.

Objective:

Achieve relative rotation of the balls around each other without dropping them.

Action Space:

The whole set of muscle [0, 1]

Observation Space:

All joints angles [-\(\pi\), \(\pi\)]

Variants:

• Easy: Swap the position of the balls myoChallengeBaodingP1-v1

• Hard: Large range of goal position and rotations. Variability in object size and frictions. myoChallengeBaodingP2-v1

• More: Sparse rewards, 3 different dense reward options to choose from

Leg Walk

A myoLeg model walks across a flat (or rough) surface. This task requires the control of 80 muscles while stabilizing the body to not fall down.

Objective:

Achieve target velocities while periodically moving your hip joints.

Variants:

• Easy: Achieve a forward velocity in the y-direction without moving in the x-direction. myoLegWalk-v0

• Hard: Achieve a forward velocity in the y-direction without moving in the x-direction on uneven terrain. myoLeg<Rough|Hilly|Stair>Walk-v0

Task and Variantions

For convenience, Myosuite has implemented a set of muscle decifiencies that currently includes: Muscle Fatigue , sarcopenia and tendon transfer alongside with the tasks.

Here are a list of the tasks currently implement in Myosuite. See Variations on how to include the additional features.

Tasks	Environment	Difficulty	Sarcopenia	Muscle Fatigue	Physical tendon transfer
Finger Joint Pose	myoFingerPoseFixed-v0	Easy	✓	✓
Finger Joint Pose	myoFingerPoseRandom-v0	Hard	✓	✓
Finger Tip Reach	myoFingerReachFixed-v0	Easy	✓	✓
Finger Tip Reach	myoFingerReachRandom-v0	Hard	✓	✓
Elbow Joint Pose	myoElbowPose1D6MRandom-v0	Hard	✓	✓
Hand Joints Pose	myoHandPoseFixed-v0	Easy	✓	✓	✓
Hand Joints Pose	myoHandPoseRandom-v0	Hard	✓	✓	✓
Hand Tips Reach	myoHandReachFixed-v0	Easy	✓	✓	✓
Hand Tips Reach	myoHandReachRandom-v0	Hard	✓	✓	✓
Hand Key Turn	myoHandKeyTurnFixed-v0	Easy	✓	✓	✓
Hand Key Turn	myoHandKeyTurnRandom-v0	Hard	✓	✓	✓
Hand Object Hold	myoHandObjHoldFixed-v0	Easy	✓	✓	✓
Hand Object Hold	myoHandObjHoldRandom-v0	Hard	✓	✓	✓
Hand Pen Twirl	myoHandPenTwirlFixed-v0	Easy	✓	✓	✓
Hand Pen Twirl	myoHandPenTwirlRandom-v0	Hard	✓	✓	✓
Die Rotation	myoChallengeDieReorientP1-v1	Easy	✓	✓	✓
Die Rotation	myoChallengeDieReorientP2-v1	Hard	✓	✓	✓
Hand Baoding Balls	myoChallengeBaodingP1-v1	Easy	✓	✓	✓
Hand Baoding Balls	myoChallengeBaodingP2-v1	Hard	✓	✓	✓
8 Obj. Rotation	myoHandReorient8-v0	Easy	✓	✓	✓
100 Obj. Rotation	myoHandReorient100-v0	Easy	✓	✓	✓
1000 Obj. Rotation	myoHandReorientID-v0	Hard	✓	✓	✓
1000 Obj. Rotation	myoHandReorientOOD-v0	Hard	✓	✓	✓
Leg walk	myoLegWalk-v0	Easy	✓	✓
Leg walk Rough Ground	myoLegRoughTerrainWalk-v0	Hard	✓	✓
Leg walk Hilly Ground	myoLegHillyTerrainWalk-v0	Hard	✓	✓
Leg walk Stair Ground	myoLegStairTerrainWalk-v0	Hard	✓	✓
Grasping & Placing	myoChallengeRelocateP1-v0	Easy	✓	✓
Chase Tag	myoChallengeChaseTagP1-v0	Easy	✓	✓

Variations

• Sarcopenia: myoSarc<Environment> e.g. myoSarcHandPoseFixed-v0

• Fatigue: myoFati<Environment> e.g. myoFatiElbowPose1D6MRandom-v0

• TTransfer / Reafferentation: myoReaf<Environment> e.g. myoReafHandPoseFixed-v0