Tasks

Arnold is trained and evaluated on 14 tasks spanning four musculoskeletal models from the MyoSuite library. Any <task_name> argument in the training, evaluation and plotting scripts must be one of the code names below.

Code names vs. paper names

The task identifiers used in the code differ from the names used in the paper. The mapping is given in the tables below — in particular reorient is Die reorient, relocate is Object relocation, and kinesis is Walk to point.

MyoElbow

Six muscles, one joint — the simplest of the four models.

Code name	Paper name	Description	Max steps	Envs
elbow_pose	Elbow pose	Rotate the elbow to point the hand at a random target location.	100	2

MyoHand

39 muscles, 23 joints. Eleven of the 14 tasks use this model — five finger-reaching tasks and six object-manipulation tasks.

Code name	Paper name	Description	Max steps	Envs
hand_thumb_reach	Thumb reach	Point the tip of the thumb at a random target location.	100	2
hand_index_reach	Index reach	Point the tip of the index finger at a random target location.	100	2
hand_middle_reach	Middle reach	Point the tip of the middle finger at a random target location.	100	2
hand_ring_reach	Ring reach	Point the tip of the ring finger at a random target location.	100	2
hand_little_reach	Little reach	Point the tip of the little finger at a random target location.	100	2
pen	Pen reorient	Rotate a pen (cylinder) to a random desired orientation.	100	2
reorient	Die reorient	Rotate a die (cube) to a random desired orientation.	150	2
baoding_p1_cw	Baoding CW	Rotate two Baoding balls clockwise. Initial phase and target rotation speed are fixed.	200	4
baoding_p1_ccw	Baoding CCW	Rotate two Baoding balls counter-clockwise. Initial phase and target rotation speed are fixed.	200	4
baoding_p2	Baoding hard	Rotate two Baoding balls. Initial phase fixed; rotation direction and target speed vary.	200	6
baoding_p2_overlap	Baoding harder	Rotate two Baoding balls. Initial phase, rotation direction and target speed all vary.	200	6

These 11 tasks are exactly the set used for the CSI analysis, which relies on all of them sharing the same 39-muscle action space.

MyoArm

63 muscles, 38 joints — the MyoHand extended with upper arm, pectoral and shoulder muscles.

Code name	Paper name	Description	Max steps	Envs
relocate	Object relocation	Grasp a dynamically generated object, lift it, and place it inside a box.	150	6

MyoLeg

80 muscles, 28 joints — a model of the human lower body.

Code name	Paper name	Description	Max steps	Envs
kinesis	Walk to point	Walk the body to a random target location without falling.	150	12

Requires the data/kinesis/ MuJoCo assets from Zenodo.

Environment allocation

The Envs column above is the number of parallel environments the paper allocates to each task during multi-task training. The allocation is deliberately imbalanced: the Baoding, Object relocation and Walk to point tasks need substantially more environment interactions, so they get more of the rollout budget.

The training scripts encode this by repeating a task name in --tasks — each occurrence gets its own --num_envs_per_task environments. With --num_envs_per_task 2, a task listed three times receives 6 parallel environments. See Training.

Full list

For copy-pasting into a --tasks argument (one occurrence each — not the imbalanced allocation):

hand_thumb_reach hand_index_reach hand_middle_reach hand_ring_reach hand_little_reach
reorient pen baoding_p1_cw baoding_p1_ccw baoding_p2 baoding_p2_overlap elbow_pose
relocate kinesis