MSK-Bench | Full-Body Musculoskeletal Motor Control

Abstract

Musculoskeletal (MSK) humanoids provide a physiologically grounded embodiment for studying full-body motor control, but their high-dimensional muscle actuation, delayed activation dynamics, and redundant muscle-tendon structures make learning substantially harder than torque-driven humanoid control. Existing MSK benchmarks remain fragmented across gait, prosthetics, dexterous hands, or challenge-specific tracks, leaving full-body muscle-actuated control insufficiently evaluated under standardized tasks, methods, and metrics. We introduce MSK-Bench, a benchmark of 22 full-body motor-control tasks organized into three progressively challenging categories: postural stabilization, common locomotor behaviors, and contact-rich environmental interaction. Under unified task protocols and robustness perturbations, MSK-Bench evaluates 5 representative control paradigms, including reward-based RL, agentic reward tuning, latent-action RL, imitation-prior control, and residual adaptation over imitation priors. Beyond task success and reward, MSK-Bench further reports robustness analysis and physiology-oriented diagnostics, including activation cost, joint smoothness, and EMG-envelope similarity. Our empirical study shows that embodiment-aware exploration and structured action representations improve task coverage in high-dimensional muscle spaces, imitation priors enhance reference-compatible stabilization and locomotion but degrade under contact-rich terrain mismatch, and residual adaptation can recover successful behaviors when fixed references fail. We further find that improved task success does not necessarily imply improved physiological agreement, highlighting the importance of evaluating task performance, robustness, and physiological behavior jointly. MSK-Bench provides a task-method-metric testbed for full-body muscle-actuated humanoid control.

Overview of MSK-Bench task families and method evaluation radar charts — **Overview of MSK-Bench (Fig. 1).** An open-source suite of **22 tasks** across stabilization, locomotion, and interaction, evaluated through **5 control paradigms** and **7 metric families**. Radar charts use eight reported axes derived from these metrics, averaged over applicable tasks in each family. EMG-envelope similarity is evaluated only where matched human reference data are available; cumulative reward is reported separately in the learning curves.

Benchmark / Framework	Embodiment		Environment		Task Design		Evaluation		Main Focus
Benchmark / Framework	Actuation (Dim.)	Full Body	3D Scene / Terrain	Contact Rich	Task Hierarchy	Tasks / Data Scale	Physio. Eval. Metrics	Control Families	Main Focus
Torque-Driven Humanoid Benchmarks
HumanoidBench [87]	Torque (19/61)	✓	✓	✓	✓	27 tasks	✗	1	Loco-manipulation
SkillBench / SkillBlender [49]	Torque (19)	✓	●	✓	✓	4 skills / 8 tasks	✗	1	Skill Blending
Mimicking-Bench [60]	Torque (19)	✓	✓	✓	✗	6 tasks	✗	1	Scene Interaction
Musculoskeletal Benchmarks
LocoMuJoCo [2]	Mixed	✓	✗	✗	●	12 envs / 27 tasks	✗	2	Locomotion Imitation
MyoSuite [12]	Muscle (≤ 80)	✗	✗	✓	●	204 tasks	✗	1	Dexterity and Agility
MyoDex [13]	Muscle (39)	✗	✗	✓	✗	14 train / 34 eval.	✗	1	Hand Manipulation
MS-Human-700 / MsGym [126]	Muscle (700)	✓	✗	●	✗	3 tasks	✓	1	MSK Locomotion
MyoChallenge 2024 [107]	Muscle (≤ 80)	●	✓	✓	●	2 tracks	✗	1	Prosthetics
MyoChallenge 2025 [15]	Muscle (≤ 80)	●	●	✓	●	2 tasks / 4 tracks	✗	1	Athletic Control
MSK-Bench (ours)	Muscle (416)	✓	✓	✓	✓	22 tasks	✓	5	Full-body Terrain MSK

Benchmark comparison. ✓/●/✗ denote full/partial/no support in the cited evaluations. We compare actuation, full-body embodiment, scene/terrain support, contact-rich interaction, task hierarchy, benchmark scale, physiological evaluation, and evaluated control families. Reference numbers follow the paper.

Evaluation Protocol

One task suite, five control paradigms, seven metric families.

All 22 tasks use MuJoCo and the same 416-muscle elastic-tendon full-body model. Success requires task completion, survival, and safety under fixed task-specific criteria. Within each task, the four full-suite baselines share observations, rewards, termination conditions, and evaluation.

Control paradigms & study scope

Reward-based RL: PPO, SAC, DepRL, and DynSyn-SAC; each is trained and evaluated on all 22 tasks.
Agentic reward tuning: focused DepRL studies with bounded reward-coefficient updates using DeepSeek-V4-Flash and GPT-6 Astra.
Latent-action RL: a focused study of an expert-trained, state-conditioned encoder–decoder; compression changes the command representation, not the executed muscle-action space.
Imitation-prior control: pretrained MuscleMimic evaluated on stand, jump, walk, run, and stairs.
Residual adaptation: a frozen imitation prior with bounded corrections and task-specific objective/timing changes on walk, run, and stairs.

Focused studies differ in task coverage and may change reward coefficients or introduce reference objectives. They are reported separately from the full-suite ranking.

What the seven metric families measure

Success Rate: successful evaluation episodes in the restored training environment, including native noise. Family and full-suite means weight tasks equally.
Cumulative Reward: task-specific evaluation return, never pooled across tasks.
Peak-Efficiency Steps: the evaluated training step attaining the highest mean return.
Perturbation Robustness: normalized trapezoidal area under success versus perturbation scale, averaged equally over tasks and action, observation, and muscle-dynamics sweeps.
Activation Cost: mean squared muscle activation, used as an effort proxy.
Joint Smoothness: mean squared angular jerk from finite differences of joint velocities, displayed logarithmically.
EMG-Envelope Similarity: mean per-muscle Pearson correlation after 101-point cycle resampling, per-muscle min–max normalization, cycle averaging, and correlation-maximizing cyclic alignment.

Action/dynamics scales: 0, 0.05, 0.10, 0.15, 0.20; observation scales: 0, 0.02, 0.05, 0.08, 0.10. Activation cost and smoothness are interpreted only for successful, behaviorally comparable policies. EMG correlations assess phase-optimized waveform shape, not absolute amplitude or timing.

Task Families

From postural regulation to contact-rich environmental interaction.

Leaderboard

Benchmark Performance and Robustness Evaluation

22 tasks across three families, comparing DepRL, SAC, PPO, and DynSyn-SAC.

Train-env. SR retains native training noise; robustness columns report separate perturbation sweeps. See the paper appendix for evaluation details. Bold SR values mark the highest training-environment success rate within each task, including ties. Activation cost and smoothness should be interpreted for successful, behaviorally comparable policies.

Stabilization

6 tasks · 4 algorithms

Scroll horizontally for all metrics and vertically for all tasks. Task and algorithm columns stay visible.

Task	Algorithm	Activation Cost	Robustness (%)															Max Steps (× 10⁷)	Train-env. SR (%) ↑	log₁₀ smooth (rad²/s⁶)
			Action					Observation					Dynamic
			Action perturbation 0	Action perturbation 0.05	Action perturbation 0.1	Action perturbation 0.15	Action perturbation 0.2	Observation perturbation 0	Observation perturbation 0.02	Observation perturbation 0.05	Observation perturbation 0.08	Observation perturbation 0.10	Dynamic perturbation 0	Dynamic perturbation 0.05	Dynamic perturbation 0.1	Dynamic perturbation 0.15	Dynamic perturbation 0.2
stand	DepRL	0.4488	94	94	90	96	98	100	12	0	0	0	98	48	28	12	4	3.74	86	3.152
	SAC	0.0484	100	100	100	100	96	100	0	0	0	0	100	100	100	100	100	7.13	100	1.875
	PPO	0.3691	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	1.97	0	5.479
	DynSyn-SAC	0.2859	52	42	56	46	36	46	46	22	30	14	40	32	40	30	30	5.44	52	2.104
powerlift	DepRL	0.3356	68	68	64	56	52	62	62	38	4	0	66	70	52	42	38	4.14	38	2.631
	SAC	0.2106	98	94	94	88	80	96	0	0	0	0	98	94	88	64	2	6.78	90	2.825
	PPO	0.3955	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	1.76	0	5.042
	DynSyn-SAC	0.3631	100	100	100	98	92	100	100	96	94	90	100	98	94	90	82	8.50	100	5.781
squat	DepRL	0.3781	44	46	38	32	30	24	56	34	62	26	46	38	12	20	10	14.02	42	2.984
	SAC	0.1184	0	0	0	4	0	0	0	0	0	0	4	0	0	0	0	4.79	0	1.708
	PPO	0.3455	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	2.01	0	5.390
	DynSyn-SAC	0.1955	48	38	32	26	22	50	44	38	36	30	48	42	32	26	18	7.00	42	6.503
sit	DepRL	0.3327	96	100	100	98	92	92	100	100	100	94	100	92	96	100	90	4.88	96	2.872
	SAC	0.1164	80	80	74	70	64	84	0	0	0	0	78	100	76	70	52	7.39	64	3.204
	PPO	0.4517	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	2.37	0	5.350
	DynSyn-SAC	0.3302	80	72	70	68	60	80	80	84	82	78	92	84	78	60	26	9.27	82	3.145
singlestand	DepRL	0.3977	14	16	28	18	10	24	30	8	0	0	12	28	6	8	12	2.10	14	2.854
	SAC	0.0849	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	4.33	0	1.954
	PPO	0.3981	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	1.84	0	5.436
	DynSyn-SAC	0.3845	24	28	24	20	18	26	22	16	0	0	22	18	8	0	0	10.00	24	2.931
balance	DepRL	0.3775	60	40	26	20	14	56	0	0	0	0	46	12	36	22	18	8.86	26	2.822
	SAC	0.1459	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	8.11	0	2.547
	PPO	0.3859	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	1.25	0	4.861
	DynSyn-SAC	0.2735	40	38	20	16	14	54	40	8	2	0	52	8	0	0	0	10.50	44	6.066

Locomotion

6 tasks · 4 algorithms

Scroll horizontally for all metrics and vertically for all tasks. Task and algorithm columns stay visible.

Task	Algorithm	Activation Cost	Robustness (%)															Max Steps (× 10⁷)	Train-env. SR (%) ↑	log₁₀ smooth (rad²/s⁶)
			Action					Observation					Dynamic
			Action perturbation 0	Action perturbation 0.05	Action perturbation 0.1	Action perturbation 0.15	Action perturbation 0.2	Observation perturbation 0	Observation perturbation 0.02	Observation perturbation 0.05	Observation perturbation 0.08	Observation perturbation 0.10	Dynamic perturbation 0	Dynamic perturbation 0.05	Dynamic perturbation 0.1	Dynamic perturbation 0.15	Dynamic perturbation 0.2
walk forward	DepRL	0.4206	28	28	32	32	44	34	32	46	24	30	40	20	12	8	6	3.32	64	3.151
	SAC	0.1427	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	9.88	0	3.262
	PPO	0.3940	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	2.42	0	5.558
	DynSyn-SAC	0.2771	82	88	68	60	48	82	82	80	76	76	82	76	76	64	8	5.00	76	3.417
run	DepRL	0.3354	30	22	32	36	42	18	18	30	32	26	32	16	12	8	2	6.72	18	3.343
	SAC	0.0940	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	2.13	0	2.854
	PPO	0.3696	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	2.16	0	5.524
	DynSyn-SAC	0.3221	96	94	84	70	68	96	94	90	86	80	96	88	2	2	0	10.00	96	3.435
walk turn	DepRL	0.3385	46	60	62	66	68	34	40	58	40	26	48	46	40	36	36	6.38	42	3.057
	SAC	0.1571	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0.99	0	2.610
	PPO	0.3747	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	1.05	0	5.331
	DynSyn-SAC	0.3112	96	88	84	78	74	94	90	90	88	84	92	24	0	0	6	8.00	86	3.248
jump	DepRL	0.3616	8	24	20	12	10	18	18	8	4	0	14	10	12	6	0	6.89	10	3.188
	SAC	0.1342	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	3.92	0	2.853
	PPO	0.3880	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	2.07	0	5.640
	DynSyn-SAC	0.2700	70	40	8	0	0	62	54	50	48	42	42	30	2	2	0	8.00	54	3.449
sidestep	DepRL	0.3888	40	38	62	56	32	34	12	2	0	0	24	22	18	10	6	4.38	28	3.349
	SAC	0.1045	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0.000032	0	3.166
	PPO	0.3729	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	4.92	0	5.408
	DynSyn-SAC	0.3630	74	66	54	42	30	74	78	74	78	54	64	72	60	40	40	5.58	76	3.223
crawl	DepRL	0.3540	100	100	100	100	98	96	100	100	100	100	100	100	98	100	100	5.12	86	3.158
	SAC	0.0338	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	2.87	0	2.516
	PPO	0.4036	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	1.34	0	5.674
	DynSyn-SAC	0.2766	100	100	100	100	100	100	100	100	100	100	100	100	100	100	100	10.00	100	3.444

Interaction with Environment

10 tasks · 4 algorithms

Scroll horizontally for all metrics and vertically for all tasks. Task and algorithm columns stay visible.

Task	Algorithm	Activation Cost	Robustness (%)															Max Steps (× 10⁷)	Train-env. SR (%) ↑	log₁₀ smooth (rad²/s⁶)
			Action					Observation					Dynamic
			Action perturbation 0	Action perturbation 0.05	Action perturbation 0.1	Action perturbation 0.15	Action perturbation 0.2	Observation perturbation 0	Observation perturbation 0.02	Observation perturbation 0.05	Observation perturbation 0.08	Observation perturbation 0.10	Dynamic perturbation 0	Dynamic perturbation 0.05	Dynamic perturbation 0.1	Dynamic perturbation 0.15	Dynamic perturbation 0.2
stair	DepRL	0.3499	52	54	38	42	40	34	38	2	0	0	32	40	24	14	4	1.42	36	2.982
	SAC	0.1993	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	6.98	0	2.459
	PPO	0.3854	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	1.57	0	5.303
	DynSyn-SAC	0.2759	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	5.92	0	2.854
catch	DepRL	0.3247	80	80	76	76	84	62	48	80	72	46	78	40	42	38	56	7.23	28	2.919
	SAC	0.1384	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	1.37	0	2.899
	PPO	0.3964	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	3.12	0	5.300
	DynSyn-SAC	0.3109	70	64	60	60	58	60	40	34	26	22	68	68	60	54	50	7.38	66	3.134
hurdle	DepRL	0.3715	48	48	56	50	40	64	44	48	28	0	62	44	40	32	20	1.26	50	3.212
	SAC	0.1852	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	4.12	0	2.816
	PPO	0.3668	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	2.46	0	5.398
	DynSyn-SAC	0.3717	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	5.50	0	7.102
slide	DepRL	0.3330	100	22	22	16	12	100	16	4	4	0	100	8	14	4	6	11.32	100	3.106
	SAC	0.0925	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	4.66	0	2.752
	PPO	0.3934	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	4.13	0	5.345
	DynSyn-SAC	0.2649	0	2	2	0	0	0	0	0	0	0	0	2	2	0	0	3.83	0	2.864
steppingstones	DepRL	0.3606	42	46	44	40	38	46	42	10	20	2	54	38	24	18	18	8.32	58	2.985
	SAC	0.1647	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	7.33	0	2.140
	PPO	0.4020	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	1.27	0	5.211
	DynSyn-SAC	0.3569	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	10.00	0	6.764
polewalk	DepRL	0.4080	22	38	16	40	20	42	26	6	0	0	30	26	20	16	16	5.14	12	2.876
	SAC	0.1504	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	6.10	0	1.724
	PPO	0.4274	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	2.87	0	5.377
	DynSyn-SAC	0.2909	98	96	98	92	90	98	0	0	0	0	96	94	94	76	16	9.00	88	3.242
reach	DepRL	0.3410	84	86	84	94	82	94	82	90	62	34	92	88	82	82	74	9.34	84	2.975
	SAC	0.1705	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	5.85	0	2.677
	PPO	0.3965	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	2.87	0	5.152
	DynSyn-SAC	0.2556	4	14	0	4	4	2	6	6	2	4	10	8	6	4	0	2.10	0	2.840
walkandsit	DepRL	0.3648	30	26	28	14	10	28	36	32	18	10	22	28	22	14	12	2.68	16	3.365
	SAC	0.1180	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	6.54	0	2.410
	PPO	0.3951	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	3.03	0	5.636
	DynSyn-SAC	0.3391	48	58	66	60	52	48	54	40	36	34	42	48	40	36	18	0.45	58	7.145
chinup	DepRL	0.3998	50	59	50	62	54	46	72	64	62	56	70	62	64	60	52	5.46	32	2.862
	SAC	0.4618	64	86	76	66	58	64	50	74	72	66	70	70	68	64	60	1.37	28	2.703
	PPO	0.3620	62	74	72	66	60	64	66	82	80	72	74	66	56	60	76	1.97	74	4.710
	DynSyn-SAC	0.2797	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	10.00	0	6.989
open door	DepRL	0.3304	96	100	100	100	98	100	94	74	62	42	100	96	96	78	66	4.71	98	3.030
	SAC	0.0962	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	1.97	0	2.836
	PPO	0.4112	0	0	0	0	0	0	0	0	0	0	0	0	0	0	0	1.77	0	5.327
	DynSyn-SAC	0.2784	100	98	98	96	92	100	100	100	98	98	100	100	14	4	0	9.50	100	6.461

Results & Analysis

Performance, robustness, and physiological behavior.

T1

Controller rankings depend on task family and metric.

DynSyn-SAC has the highest full-suite mean success rate (52.00%), driven primarily by locomotion. DepRL performs better on interaction tasks, has nonzero observed success on 22/22 tasks, and scores higher on action and dynamics robustness; DynSyn-SAC scores higher on observation robustness. These are separately trained task-specific policies. Focused studies reveal trade-offs in reward adaptation and action representation, rather than a uniformly superior design.

Full-suite point estimates · Table II
Metric ↑	DynSyn-SAC	DepRL	SAC	PPO
Stabilization SR (%) · 6 tasks	57.33	50.33	42.33	0.00
Locomotion SR (%) · 6 tasks	81.33	41.33	0.00	0.00
Interaction SR (%) · 10 tasks	31.20	51.40	2.80	7.40
Mean SR (%) · 22 tasks	52.00	48.36	12.82	3.36
Action robustness	46.66	53.06	15.36	3.10
Observation robustness	43.37	38.10	4.26	3.40
Dynamics robustness	34.75	39.84	14.50	2.92
Mean robustness	41.59	43.67	11.37	3.14
Coverage · SR > 0	16/22	22/22	4/22	1/22

SR retains training noise. Means weight tasks equally. Robustness areas are normalized to 0–100, with the average weighting perturbation types equally. Coverage counts tasks with nonzero observed success, not universal success. Bold indicates the best point estimate in each row; these are not claims of statistical significance.

Task-specific reward-learning curves for 22 tasks, comparing DepRL, DynSyn-SAC, SAC, and PPO — **Task-specific reward-learning curves for 22 tasks (Fig. 2).** DynSyn-SAC attains higher returns than DepRL on jump and hurdle over the displayed overlapping ranges, while DepRL leads on walk-turn and catch; SAC remains competitive on powerlift. Shading shows across-run variation. The full-suite table separately reports success, nonzero-success coverage, and robustness over all 22 tasks.

Agentic reward tuning reward weights and robustness perturbation plots — **Reward adaptation and robustness (Fig. 3).** In the original Agentic-DepRL setting, coefficients are updated every 20,000 steps and each nonzero-weight change is limited to 20%; unnormalized coefficient magnitudes still grow. Fixed-reward DepRL maintains higher survival under the tested action and observation perturbations, while agentic tuning leads at larger dynamics-perturbation scales. Both share learner, architecture, and budget. Each point averages 50 held-out seeds; survival is capped at 1,000 steps.

Matched reasoner comparison · Table III
Setting	Survival (steps) ↑	Tracking RMSE ↓	Activation ↓
Fixed reward	417	0.4446	0.4769
DeepSeek	246.6	0.4355	0.3999
GPT-6 Astra	360	0.4324	0.3879

DeepSeek and GPT-6 Astra use the same DepRL learner, architecture, training budget, and evaluation, with an identical reward-coefficient constraint ‖w‖₁ = 40. GPT-6 Astra has better point estimates than DeepSeek on all three measures, but shorter survival than fixed-reward DepRL despite lower RMSE and activation.

Direct muscle-control diagnostic · Table IV
Task	Duration (s) ↑	Divergence (s) ↑	Tracking RMSE ↓	Outcome
Walk	0.75	0.53	0.158	Fall
Run	2.02	0.94	0.653	Fall
Stairs	0.80	N/A	0.173	Fall

GPT-6 Astra directly outputs 354-dimensional muscle excitations at 5 Hz in a 100-Hz simulation, holding each action for 0.2 s, without training or a learned low-level controller. One seed-0 rollout per task uses a 3-s horizon. Divergence denotes the first sustained tracking divergence; N/A means the fixed criterion was not met. These outcomes characterize only this tested configuration.

Latent action compression reward and robustness plots — **128D**latent dimension

T2

Task-specific tests characterize a stair-control failure.

Under the tested dynamics perturbations, MuscleMimic generally has higher measured success than DepRL on stand, jump, walk, and run. On stairs, MuscleMimic has no observed successes at any tested scale, including zero dynamics perturbation. Tracking errors characterize divergence from the reference without establishing a timing mismatch or identifying the cause of failure. This task-specific limitation motivates the combined residual adaptation study below.

Robustness comparison and tracking error analysis between imitation prior and DepRL — **Task-dependent dynamics robustness and stair-tracking errors (Fig. 5).** (a–e) Success rates of MuscleMimic/MM (PPO) and DepRL on stand, jump, walk, run, and stairs. MuscleMimic has no observed stair successes, including at zero perturbation. (f) From step 150 to 300, absolute-site error rises from 0.35 to 1.31 and root-XYZ error from 0.25 to 0.77. These errors describe reference divergence; they do not establish its cause. Both methods share success conditions and perturbation settings within each task. This focused protocol differs from the full-suite benchmark; its values are excluded from aggregate robustness scores.

T3

Task success does not characterize muscle use.

Walking retains 100% success after adaptation while selected muscle correlations improve; running improves in both success and mean EMG correlation. Stairs recovers from 0% to 100% success, but mean correlation changes only from 0.51 to 0.52 and remains below DepRL’s 0.56. Muscle-wise EMG comparisons, task-dependent joint jerk, load-specific recruitment, and anatomical resolution reveal differences that aggregate task scores do not capture.

Walking EMG-envelope similarity across policies — **Walking EMG-envelope comparison (Fig. 6).** Human recordings from Gait120 versus simulated activations for reward-based RL, latent-action RL, agentic tuning, and imitation-prior control. Signals are resampled to 101 gait-cycle points, min–max normalized per muscle, cycle-averaged, and cyclically aligned to maximize Pearson correlation. Correlations assess phase-optimized waveform shape, not absolute amplitude or timing. Agreement varies by muscle: latent-action RL improves biceps femoris correlation (0.95 vs. 0.84) but lowers vastus lateralis correlation (0.21 vs. 0.73) relative to reward-based RL. Walking and stair references use Gait120; running uses the 3-m/s recordings of Van Hooren and Meijer.

Task-wise joint smoothness across benchmark tasks — **Task-wise joint smoothness (Fig. 7).** Bars report log10 mean squared angular jerk computed from finite differences of joint velocities. Levels depend on the task, with overlapping ranges across task families. This characterizes movement fluctuations, not a cross-task ranking of success, control quality, or human similarity.

Load-dependent muscle-family activation during powerlift — **0.05kg**barbell weight

Effect of anatomical resolution on powerlift muscle recruitment — **416-muscle model**shared muscle families

Residual Adaptation

Task recovery and EMG agreement must be assessed separately.

A case study connects stair-failure diagnosis to multi-metric validation. A frozen MuscleMimic prior supplies base actions and a trainable three-layer MLP supplies bounded corrections. The pipeline also changes task objectives and reference timing; its results do not isolate the effect of residual actions alone.

Combined adaptation pipeline

a_full = clip(a_base + α clip(a_res, −1, 1), −1, 1)

Stairs: reference progression is normal below 0.4 m planar root error and slows by 50% otherwise. Vertical root displacement is excluded from tracking; progress targets the reference state 1 s ahead.
Walking: residual authority is 0.6 in swing and 0.2 in stance, with activation magnitude and temporal-change penalties.
Running: flight-phase tracking penalties are halved, toe-off is rewarded, and vertical ground-reaction forces above 2.5 times body weight receive a soft penalty.

A small L2 residual penalty (c_res = −0.02) replaces the generic action-rate and activation-cost terms. The 354-dimensional actions in the controller diagram exclude hand actuation.

MuscleMimic

Stairs

Run

Walk

Residual adaptation

Stairs

Run

Walk

MuscleMimic and the combined residual adaptation pipeline. The top row shows imitation-prior rollouts for stairs, running, and walking, with adapted rollouts below. Videos illustrate behavior; task success and EMG-envelope agreement are reported separately in Table V.

Residual adaptation results · Table V
Task	MM SR (%) ↑	Residual SR (%) ↑	MM EMG r ↑	DepRL EMG r ↑	Residual EMG r ↑
Walking	100	100	0.71 / 0.69 / 0.63	—	0.75 / 0.80 / 0.67
Running	90	100	0.43	0.70	0.77
Stairs	0	100	0.51	0.56	0.52

MM = MuscleMimic. SR retains training noise. Walking reports soleus medialis / biceps femoris / semitendinosus correlations; running reports an 11-channel phase-aligned mean, and stairs reports a 12-muscle phase-aligned mean. The dash means no DepRL walking value is reported in this table. Bold marks the best point estimate in each comparison. Task recovery does not imply uniformly better human EMG agreement.

Additional diagnostics: stair height rises from 0.94 to 1.21 m and tracking error falls from 0.1968 to 0.1079 rad. Running tracking error falls from 0.2267 to 0.1338 rad and peak ground-reaction force from 6.40 to 6.07 times body weight.

Limitations & Future Work

Scope of the benchmark and next steps.

The reported results describe the evaluated tasks and perturbations; they do not establish statistical dominance or direct transfer to unseen embodiments. The shared 416-muscle humanoid controls for body design, but conclusions may not generalize to other anatomical models.

Human-reference EMG evaluation covers three tasks with task-dependent lower-limb signals: three selected muscles for walking in Table V, 11 independent channels for running, and 12 matched muscles for stairs. The correlations measure phase-optimized activation-envelope shape rather than complete biological similarity. Residual adaptation changes actions, rewards, and reference timing together, so its gains characterize the complete pipeline.

Future extensions include more dexterous manipulation tasks, longer-horizon behaviors, real-world validation, and broader physiological measurements.

Citation

BibTeX

@misc{ou2026mskbenchbenchmarkingfullbodymusculoskeletal,
      title={MSK-Bench: Benchmarking Full-Body Musculoskeletal Motor Control Across Tasks, Control Paradigms, and Physiological Metrics},
      author={Mengtao Ou and Zongzheng Zhang and Zhenghao Xiao and Yixuan Pan and Ziwen Zhuang and Hang Zhao and Hongyang Li and Yanan Sui and Libin Liu and Hao Zhao},
      year={2026},
      eprint={2609.26872},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2609.26872},
}

MSK-Bench: Benchmarking Full-Body Musculoskeletal Motor Control Across Tasks, Control Paradigms, and Physiological Metrics MSK-Bench: Benchmarking Full-Body Musculoskeletal Motor Control Across Tasks, Control Paradigms, and Physiological Metrics

One task suite, five control paradigms, seven metric families.

Control paradigms & study scope

What the seven metric families measure

From postural regulation to contact-rich environmental interaction.

Benchmark Performance and Robustness Evaluation

Stabilization

Locomotion

Interaction with Environment

Performance, robustness, and physiological behavior.

Controller rankings depend on task family and metric.

Task-specific tests characterize a stair-control failure.

Task success does not characterize muscle use.

Task recovery and EMG agreement must be assessed separately.

Combined adaptation pipeline

Scope of the benchmark and next steps.

BibTeX