muscle-supermimic checkpoints

Policies trained with muscle-supermimic (SuperMimic: world-model BPTT motion tracking, on mjlab / MuJoCo-Warp) for muscle-driven humanoids. Private: see Data and licenses.

The checkpoints of iteration 16000 of both runs below are the ones the project page's live demo runs in the browser.

The project page's runs

Both runs track MuscleMimic's KIT training list (972 clips) with MuscleMimic's MDP (100 Hz, MimicReward, deviation terminations) and the same settings: Euler, 100 solver / 50 line-search iterations, timestep 2 ms, decimation 5; 4096 envs, horizon 32, pi_batch 4096, lr 2e-4, excitation noise N(0, 0.1), state_grad_clip 100, seed 42; one RTX 5090 each.

mshuman700/kit_mm_euler_sgc100/model_{010000,016000}.pt

MS-Human-700 with locked hands (694 muscles), on the KIT clips retargeted from MuscleMimic's MyoFullBody clips with ms-retarget-myofullbody (training list kit_kinesis_train.txt of that data).

task	Mjlab-SuperMimic-Flat-MS-Human-700-Locked-Hands
run	2026-10-26_01-53-49_kit_mm_euler_sgc100, checkpoints every 500 iterations (these two kept here)

iteration	training time	episode length	normalized episode return	108 held-out clips, MuJoCo-Warp	same, CPU
10000	6.2 h	276 steps	0.896	0.898, all tracked	0.898, 0 failed
16000	10.1 h	266 steps	0.912	0.911, 104 tracked	0.918, 3 failed

myofullbody/kit_n4096_euler_sgc100/model_{010000,016000}.pt

MyoFullBody (musclemimic-models==1.0.6, fingers disabled: 31 joint DoF, 354 muscles, self-contacts kept) on MuscleMimic's retargeted clips (data/musclemimic/kit_kinesis_train.txt).

task	Mjlab-SuperMimic-Flat-MyoFullBody
run	2026-10-26_06-01-53_kit_n4096_euler_sgc100, checkpoints every 2000 iterations (these two kept here)

iteration	training time	episode length	normalized episode return	108 held-out clips, MuJoCo-Warp	same, CPU
10000	6.0 h	291 steps	0.747	0.828, 101 tracked	0.828, 7 failed
16000	9.6 h	271 steps	0.861	0.903, all tracked	not run

The held-out clips are MuscleMimic's KIT test list (kit_kinesis_test.txt), each played once from its first frame without exploration noise; episode length and return are the training log's. Under C MuJoCo, as the browser demo runs them (ms-export-web --reference), the checkpoints of iteration 16000 score 0.914 (MS-Human-700, 3 failed) and 0.899 (MyoFullBody, 1 failed).

They load and evaluate with muscle-supermimic master 913d0a5:

MS_HUMAN_700_ROOT=<MS-Human-700> ms-eval --task Mjlab-SuperMimic-Flat-MS-Human-700-Locked-Hands \
    --checkpoint mshuman700/kit_mm_euler_sgc100/model_016000.pt \
    --motion-file data/mshuman700_kit/kit_kinesis_test.txt --device cuda:0 [--record rollout.npz]
ms-eval --task Mjlab-SuperMimic-Flat-MyoFullBody --checkpoint myofullbody/kit_n4096_euler_sgc100/model_016000.pt \
    --motion-file data/musclemimic/kit_kinesis_test.txt --device cuda:0 [--record rollout.npz]
ms-render rollout.npz videos/

myofullbody/kit_n4096_euler/model_012000.pt

An earlier MyoFullBody run, without state_grad_clip. MyoFullBody (musclemimic-models==1.0.6, fingers disabled: 31 joint DoF, 354 muscles, self-contacts kept) tracking MuscleMimic's KIT training list (972 clips, data/musclemimic/kit_kinesis_train.txt), MuscleMimic's MDP (100 Hz, MimicReward, deviation terminations).

task	Mjlab-SuperMimic-Flat-MyoFullBody
simulation	Euler, 100 solver / 50 line-search iterations, timestep 2 ms, decimation 5
training	4096 envs, horizon 32, pi_batch 4096, lr 2e-4, excitation noise N(0, 0.1), seed 42
iteration	12000 (10.7 h on one RTX 5090, 2026-10-25)
training metrics at 12000	episode length 268 steps, failure terminations 7 %, normalized episode return 0.710

The run's policy gradient later exploded and its tracking collapsed (from iteration ~12700); this is the last checkpoint before that. params/ holds the run's env and agent configs, train_log.csv the whole run's log.

Evaluation on 3 training clips (359/turn_left05, 11/WalkingStraightForwards05, 7/WalkingStraightBackwards04), whole clips, no failure on any:

	normalized return
deterministic, code of the run	0.787
deterministic, muscle-supermimic master a1f91ba	0.793
excitation noise N(0, 0.1), seeds 0 / 1 / 2	0.782 / 0.780 / 0.779

The run used that day's working tree (between PR #3 and PR #5); the checkpoint loads and evaluates with master a1f91ba:

ms-eval --task Mjlab-SuperMimic-Flat-MyoFullBody --checkpoint model_012000.pt \
    --motion-file data/musclemimic/kit_kinesis_train.txt --device cpu [--record rollout.npz]
ms-render rollout.npz videos/

Data and licenses

• Models: MS-Human-700 (Apache-2.0). MyoFullBody from musclemimic-models (Apache-2.0); its arm and hand derive from MoBL-ARMS, which is non-commercial only.
• Motions: AMASS KIT clips retargeted to MyoFullBody by MuscleMimic (amathislab/musclemimic-retargeted), and from those onto MS-Human-700 with ms-retarget-myofullbody; AMASS's license allows non-commercial research use only and no redistribution of the data.