Literature
Display
← All papers

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

Mengtao Ou · Zongzheng Zhang · Zhenghao Xiao · Yixuan Pan · Ziwen Zhuang · Hang Zhao · Hongyang Li · Yanan Sui · Libin Liu · Hao Zhao

Preprint posted · arXiv (preprint) ↗

Robotics12 min text · CondensedOriginal paper ↗
Contents

Overview

This edition follows the 35-page arXiv v1 dated 22 September 2026, retaining all 15 figures, 18 tables, 44 displayed equations and 127 references. The common benchmark has 22 tasks; only four reward-based baselines run the complete suite. Agentic tuning, action filtering, imitation and residual adaptation are focused studies with their own protocols. Success, robustness and agreement with human muscle signals answer different questions. Two figures use native PDF vector artwork and seven use larger original PDF rasters to restore detail lost in HTML copies.

Refer to caption
Fig. 1: Overview of the Musculoskeletal Benchmark (MSK-Bench). (a) Benchmark Components: An open-source suite featuring 22 tasks across stabilization, locomotion, and interaction, evaluated under 5 control paradigms and 7 metric families. (b) Method Evaluation: Radar charts compare control methods across the 3 task families using 8 reported axes derived from the benchmark metrics. Each value represents the average performance across all applicable tasks within the respective category. EMG-envelope similarity is evaluated only where matched human reference data are available, while cumulative reward is reported separately in Fig. 2.

At a glance

MSK-Bench standardizes full-body muscle control around one 416-muscle humanoid, task-specific success criteria and separate behavioral/physiological diagnostics. Structured exploration and action representations improve coverage, while motion priors work unevenly across tasks. Residual adaptation recovers stair navigation in the tested setting, but the corresponding human-signal correlation changes little. The evidence therefore favors evaluating task achievement and physiological agreement jointly rather than treating either as a substitute for the other.

I Introduction

Muscle excitation produces motion through delayed activation, contraction and tendon transmission, creating redundant control channels unlike directly commanded joint torques. Existing musculoskeletal benchmarks cover fragments of this problem. MSK-Bench supplies a shared model, task definitions, perturbation sweeps and interpretable diagnostics so that a controller’s strengths and failures can be compared under explicit conditions.

II Related Work

The benchmark connects torque-controlled humanoid learning, musculoskeletal simulation, imitation priors and human locomotion datasets. PPO and SAC provide generic baselines; DepRL adds structured embodied exploration, while DynSyn learns a lower-dimensional synergistic action representation [PPO], [SAC], [DepRL], [DynSyn]. Table I contrasts existing task coverage and physiology-related evaluation. Human EMG data provide limited matched references, not a comprehensive biological ground truth.

TABLE I: Benchmark comparison. ✓/∙\bullet/✗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.
Benchmark / Framework Embodiment Environment Task Design Evaluation Main Focus
Actuation (Dim.) Full Body 3D Scene / Terrain Contact Rich Task Hierarchy Tasks / Data Scale Physio. Eval. Metrics Control Families
Torque-Driven Humanoid Benchmarks
HumanoidBench [88] Torque (19/61) ✓ ✓ ✓ ✓ 27 tasks ✗ 1 Loco-manipulation
SkillBench / SkillBlender [50] Torque (19) ✓ ∙\bullet ✓ ✓ 4 skills / 8 tasks ✗ 1 Skill Blending
Mimicking-Bench [61] Torque (19) ✓ ✓ ✓ ✗ 6 tasks ✗ 1 Scene Interaction
Musculoskeletal Benchmarks
LocoMuJoCo [2] Mixed ✓ ✗ ✗ ∙\bullet 12 envs / 27 tasks ✗ 2 Locomotion Imitation
MyoSuite [15] Muscle (≤80\leq 80) ✗ ✗ ✓ ∙\bullet 204 tasks ✗ 1 Dexterity and Agility
MyoDex [12] Muscle (39) ✗ ✗ ✓ ✗ 14 train / 34 eval. ✗ 1 Hand Manipulation
MS-Human-700 / MsGym [127] Muscle (700) ✓ ✗ ∙\bullet ✗ 3 tasks ✓ 1 MSK Locomotion
MyoChallenge 2024 [108] Muscle (≤80\leq 80) ∙\bullet ✓ ✓ ∙\bullet 2 tracks ✗ 1 Prosthetics
MyoChallenge 2025 [14] Muscle (≤80\leq 80) ∙\bullet ∙\bullet ✓ ∙\bullet 2 tasks / 4 tracks ✗ 1 Athletic Control
MSK-Bench (ours) Muscle (416) ✓ ✓ ✓ ✓ 22 tasks ✓ 5 Full-body Terrain MSK

III MSK-Bench

The suite has six stabilization tasks, six locomotion tasks and ten terrain/object-interaction tasks. Each uses the same 416-muscle elastic-tendon body with shared proprioceptive/muscle state plus task-specific information. Policies are trained separately per task. PPO, SAC, DepRL and DynSyn-SAC share task observations, rewards, termination and evaluation within the full-suite comparison; focused studies must be read separately.

Full-suite train-environment success averages 52.00% for DynSyn-SAC, 48.36% for DepRL, 12.82% for SAC and 3.36% for PPO. DynSyn-SAC is strongest in locomotion (81.33%), whereas DepRL is strongest in interaction (51.40%) and has nonzero success on all 22 tasks versus 16 for DynSyn-SAC. The average normalized perturbation-area score reverses their ranking: 43.67% versus 41.59%. Coverage means at least one successful outcome, not reliable mastery. Native training noise remains in the success evaluation, so these rates need not equal the separately configured zero-perturbation sweep points.

Agentic reward updates show trade-offs rather than a uniform improvement. The original unnormalized study permits each nonzero weight to change by at most 20% every 20,000 steps, which still allows cumulative growth. The matched reasoner study instead fixes the weight L1 norm at 40: GPT-6 Astra outperforms DeepSeek-V4-Flash on the reported survival, tracking and activation point estimates, but both survive fewer steps than fixed weights. Direct GPT-6 control is a separate diagnostic: one seed-zero rollout per task, 354 action values at 5 Hz held by a 100-Hz controller, no learned low-level policy, and a three-second maximum. Walk, run and stairs fall at 0.75, 2.02 and 0.80 seconds; these outcomes describe this interface and protocol.

A 128-dimensional action representation improves selected learning/perturbation results over 416 actions; stronger compression degrades them. MuscleMimic exceeds DepRL on several reference-compatible tasks, but fails the focused stair test even without added perturbation. Its five-task comparison uses matched conditions within that study, including a less stringent Jump criterion than the full suite. Increasing stair-tracking error is a diagnostic, not an isolated causal explanation.

Activation cost is a mean-square effort proxy, joint smoothness is based on angular jerk, and EMG correlation compares normalized, optimally phase-shifted envelope shapes. None alone measures biological realism. Walking results show muscle-level reversals between controllers. Powerlift recruitment changes with load, but trunk/back activation masses of 30.1, 50.4 and 46.0 at 0.05, 5 and 10 kg are not monotonic. The 700-muscle anatomy adds recruitment pathways for analysis without establishing greater human similarity.

Refer to caption
Fig. 2: Task-specific reward-learning curves for 22 tasks. Four reward-based algorithms show task-dependent advantages. Shading shows across-run variation; Table II separately reports success, robustness, and nonzero-success coverage over all 22 tasks.
Equations 1–2
𝐮t\displaystyle\mathbf{u}_{t} ∼πθp(⋅∣ot,ct),\displaystyle\sim\pi_{\theta}^{p}(\cdot\mid o_{t},c_{t}), (1)
𝐚t\displaystyle\mathbf{a}_{t} =Fp​(ot,𝐮t,ct)∈[−1,1]M,\displaystyle=F_{p}(o_{t},\mathbf{u}_{t},c_{t})\in[-1,1]^{M},
Jp​(θ,𝐰p,k)\displaystyle J_{p}(\theta;\mathbf{w}_{p,k}) =𝔼τ∼P,πθp,Fp​[∑t=0T−1γt​r~p,t],\displaystyle=\mathbb{E}_{\tau\sim P,\pi_{\theta}^{p},F_{p}}\!\left[\sum_{t=0}^{T-1}\gamma^{t}\widetilde{r}_{p,t}\right],
r~p,t\displaystyle\widetilde{r}_{p,t} =𝐰p,k⊤​𝐟p,t−λp,k​Ωp,t+βp​ℋp,t.\displaystyle=\mathbf{w}_{p,k}^{\top}\mathbf{f}_{p,t}-\lambda_{p,k}\Omega_{p,t}+\beta_{p}\mathcal{H}_{p,t}.
𝐚tfull=clip⁡(𝐚tbase+α​clip⁡(𝐚tres,−1,1),−1,1).\mathbf{a}_{t}^{\mathrm{full}}=\operatorname{clip}\!\left(\mathbf{a}_{t}^{\mathrm{base}}+\alpha\operatorname{clip}(\mathbf{a}_{t}^{\mathrm{res}},-1,1),-1,1\right). (2)
TABLE II: Full-suite point estimates. SR retains training noise. Means weight tasks equally. Robustness areas are normalized to 0–100; Avg. weights the three perturbation types equally. Coverage: SR >0>0. Bold: row best.
Group Metric DynSyn-SAC DepRL SAC PPO
SR (%) ↑\uparrow Stabilization (6) 57.33 50.33 42.33 0.00
Locomotion (6) 81.33 41.33 0.00 0.00
Interaction (10) 31.20 51.40 2.80 7.40
Average (22) 52.00 48.36 12.82 3.36
Robust. ↑\uparrow Action 46.66 53.06 15.36 3.10
Observation 43.37 38.10 4.26 3.40
Dynamics 34.75 39.84 14.50 2.92
Average 41.59 43.67 11.37 3.14
Coverage ↑\uparrow SR >0>0 16/22 22/22 4/22 1/22
Refer to caption
Fig. 3: Reward adaptation and robustness. (a) Coefficient updates; (b–d) survival under perturbations. Agentic-DepRL and fixed-reward DepRL share learner, architecture, and training budget. Each point averages 50 held-out seeds; survival is capped at 1000 steps.
Refer to caption
Fig. 4: Action compression. Learning and perturbation tests for 416 actions and 128-, 64-, 32-, and 16-dimensional representations. The focused study improves at 128 dimensions but degrades under stronger compression; it is not a full-suite comparison.
TABLE III: Matched reasoner comparison.
Setting Survival (steps) ↑\uparrow Tracking RMSE ↓\downarrow Activation ↓\downarrow
Fixed reward 417 0.4446 0.4769
DeepSeek 246.6 0.4355 0.3999
GPT-6 Astra 360 0.4324 0.3879
TABLE IV: GPT-6 direct muscle-control diagnostic. Div.: first sustained tracking divergence; N/A indicates that the fixed divergence criterion was not met.
Task Duration (s) ↑\uparrow Div. (s) ↑\uparrow Tracking RMSE ↓\downarrow 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
Refer to caption
Fig. 5: Task-dependent performance under dynamics perturbations and stair-tracking errors. (a–e) Matched-protocol success rates of MuscleMimic/MM (PPO) and DepRL. Within each task, both methods use identical success conditions and perturbation settings. This focused cross-paradigm protocol is distinct from the full-suite benchmark protocol reported in the robustness tables; its values are not directly comparable with those tables and are excluded from aggregate robustness scores. MuscleMimic has no observed stair successes at any tested scale, including zero dynamics perturbation. (f) Reported tracking-error components on stairs characterize deviation from the reference, without identifying a timing mismatch or the cause of task failure.
Refer to caption
Fig. 6: Walking EMG-envelope comparison. Human recordings [8] versus simulated activations. Normalized, phase-optimized correlations (rR,rL,rA,rIr_{\mathrm{R}},r_{\mathrm{L}},r_{\mathrm{A}},r_{\mathrm{I}}) correspond to reward-based RL, latent-action RL, agentic tuning, and imitation-prior control.
Refer to caption
Fig. 7: Task-wise joint smoothness. Bars report log10\log_{10} mean squared angular jerk. Stabilization shows the lowest measured jerk, locomotion the highest, and interaction intermediate values. These characterize motion fluctuations, not success or human similarity.
Refer to caption
Fig. 8: Load-dependent recruitment in powerlift. Muscle-family activation is computed by summing the time-averaged activations of its constituent actuators. Bars report this activation mass under different barbell weights; the metric does not represent instantaneous activation or agreement with human recordings.
Refer to caption
Fig. 9: Anatomical resolution and powerlift recruitment. Activation mass for the 416- and 700-muscle models. The latter adds neck, deep-trunk, and deep-scapular groups, expanding available recruitment pathways; this does not establish greater human similarity.

IV Residual Adaptation Case Study

A frozen MuscleMimic prior receives a trainable residual correction together with task-specific reward and timing changes. Stairs ignore vertical root mismatch, slow reference advancement when planar error exceeds 0.4 m and reward progress toward a reference one second ahead. Walking grants more correction during swing and penalizes activation magnitude/change. Running relaxes tracking during flight, rewards propulsion and softly penalizes vertical ground-reaction force above 2.5 body weights. This is a combined adaptation pipeline, not an ablation of residual actions alone.

Table V reports walking success unchanged at 100%, running increasing from 90% to 100%, and stairs from 0% to 100%. Selected walking correlations improve; mean running correlation increases from 0.43 to 0.77. Stair correlation changes only from 0.51 to 0.52 and remains below DepRL’s 0.56. Stair height and tracking improve, while running peak ground force falls from 6.40 to 6.07 body weights. Recovering a task therefore does not guarantee a matching physiological gain.

Refer to caption
Fig. 10: Residual architecture and motions. Frozen prior plus trainable correction; blue skeletons are references. The original lower-branch label “ref action” denotes the residual correction in our equation. The 354d labels exclude hand actuation. Rewards and reference timing also change.
TABLE V: Residual adaptation results. MM: MuscleMimic. SR retains training noise. EMG rr reports SoleusMed./Bic.Fem./Semitend. for walking, an 11-channel phase-aligned mean for running, and a 12-muscle phase-aligned mean for stairs. Bold: best in each comparison.
SR (%) ↑\uparrow EMG rr ↑\uparrow
Task MM Residual MM DepRL Residual
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

V Limitations and Future Work

Results are point estimates on a fixed simulated morphology and do not establish statistical dominance or transfer to other bodies. EMG validation covers only walking, running and stairs with different matched muscle sets. Residual learning, objective changes and reference pacing are combined, so their individual contributions are unresolved. Dexterity, longer compositions, broader physiological measurements and physical hardware remain outside the demonstrated scope.

VI Conclusion

MSK-Bench makes the task–method–metric relationship explicit: structured exploration helps some high-dimensional control problems, priors help reference-compatible motions, and adaptation can recover a failure while leaving physiological agreement largely unchanged. Its contribution is a reproducible comparison framework with diagnostic detail, not a universally best controller.

Appendix A Appendix

The appendix gives exact task rewards and termination rules, code-level success criteria, perturbation definitions, training defaults, controller interfaces, recruitment grouping and adaptation details. These tables and formulas are retained because they determine what the headline results mean.

A-A Task Specification

The task definitions use joint states, body positions/velocities, center-of-mass and foot geometry, torso/chest/pelvis directions, phase, terrain grids and obstacle features. Leg extension helps distinguish standing from kneeling shortcuts. Reward terms combine task progress, posture/contact constraints, survival, activation cost and penalties. Tables VI–VIII preserve every reward formula/weight; Tables IX–XI preserve initialization, task observations and termination. Figures 11–13 retain the full visual task descriptions. These are designed benchmark objectives, not evidence that learned recruitment reproduces the biomechanics motivating them.

Refer to caption
Fig. 11: Illustration of six representative whole-body stabilization tasks. (a) squat: The control policy must accurately track the dynamic vertical trajectories of the hip and knee based on a phase clock ϕ\phi, while maintaining a strict center-of-mass balance lock between the two feet. A specially designed “landmine penalty” is introduced to immediately eliminate abnormal strategies that attempt to exploit the height-tracking reward by kneeling on both knees. (b) stand: The robot is required to maintain an absolutely static upright posture from an initial neutral pose. The key challenge is that the system must use high muscle stiffness to lock the joints, while even slight center-of-mass drift or foot sliding is severely penalized. The reward function is strongly biased toward absolute stillness and posture maintenance. (c) powerlift: To sustain the extreme load, the reward function explicitly encourages a biomechanically plausible forward-leaning hip-hinge posture through the RhingeR_{\mathrm{hinge}} reward, thereby effectively recruiting the powerful posterior-chain muscle groups. Once dangerous lumbar collapse is detected, the episode is immediately terminated to ensure structural safety. (d) Singlelegstand: The robot must lift one leg into a “flamingo” posture and strictly align the center-of-mass projection with the support foot. The task severely penalizes lateral trunk lean, and after an initial physical tolerance period of 1010 steps, the episode is immediately terminated once the lifted foot touches the ground. (e) sit: The robot is required to approach a chair and execute a precise sitting transition. To prevent cheating behaviors such as stiff-legged falling, the task enforces active knee flexion, requires the feet to be retracted beneath the support area, and introduces a “soft-landing” reward to limit pelvic descent velocity, ensuring a smooth motion. (f) balance: The robot is required to maintain dynamic balance on an unconstrained wobble board placed on a sphere. To overcome the common failure mode of falling backward, the task explicitly enforces a forward-leaning “fighting defensive stance”, with a target pitch angle of −0.15​rad-0.15\,\mathrm{rad}, and requires precise center-of-mass control to keep the balance board level.
Refer to caption
Fig. 12: Visualization of representative locomotion behaviors generated by the humanoid robot. Each subfigure shows a sequence of motion frames, with red and blue skeletons indicating the two legs and body segments over time. (a) walk turn: The robot moves while interacting with spherical obstacles. (b) walk forward: The robot demonstrates a stable periodic gait with alternating leg support. (c) sidestep: The robot demonstrates sustained rhythmic whole-body coordination. (d) run: The robot runs forward with alternating leg motion while maintaining dynamic balance. (e) crawl: The robot performs low-posture locomotion with the body close to the floor. (f) Jump: The robot performs a dynamic aerial transition while maintaining body balance.
Refer to caption
Fig. 13: Illustration of ten representative terrain and object interaction tasks. (a) walkandsit: The robot is required to approach a chair backward and sit down. Instead of using a conventional collision-avoidance penalty, the task rewards the robot for confidently utilizing downward inertia to complete the sitting motion and for comfortably leaning back against the chair after sitting, thereby reproducing human-like sitting biomechanics. (b) stepstones: A magnetic attraction reward is used to guide the foot toward the target landing positions, while strict penalties on trunk lateral bending, yaw deviation, and excessive knee flexion encourage an upright and stable traversal posture. (c) slide: The robot is required to adaptively lean forward to counteract inertia during slope transitions. A split-prevention penalty is introduced to terminate the episode once the lateral separation of the legs exceeds the biomechanical limit. (d) stairs: The robot is required to climb stairs using look-ahead terrain perception from a terrain radar. When upcoming steps are detected, an active leg-lifting reward is activated to encourage anticipatory step planning. (e) reach: The robot is required to rapidly and explosively reach randomly reset spatial targets with its dominant hand while maintaining a slightly crouched defensive posture. Since the target height varies, the task strictly constrains knee height to prevent cheating strategies such as kneeling. (f) hurdle: The task enforces a complete aerial flight phase and introduces a strict anti-flamingo-jump penalty to prevent abnormal strategies in which the robot crosses the hurdle by curling and dragging a single leg. (g) open door: The robot is required to approach, push open, and pass through a hinged door. A global attraction field guides the hand toward the door handle, while fall-prevention penalties discourage shortcut behaviors such as headbutting the door or diving through it. (h) chinup: The robot is required to pull its head above the horizontal bar from a hanging posture. The task rewards the activation of major pulling muscle groups such as the latissimus dorsi and biceps, while imposing strict penalties on body swinging and leg-kipping, forcing the motion to be completed through pure muscular strength. (i) catch: The robot is required to coordinate both arms to catch a flying cube thrown toward the chest while maintaining an upright posture. The key objective is to track the distance between both hands and the target, and a large sparse reward is given when the object is successfully caught and stabilized in mid-air. (j) polewalk: The robot uses terrain radar to navigate through a field of poles while avoiding obstacles. In accordance with the task specification and collision penalty, head or torso contact with a pole terminates the episode.
TABLE VI: Reward Functions for All 22 Musculoskeletal Benchmark Tasks (Part 1: Stabilization Series). Note: Action regularization (RefficiencyR_{\text{efficiency}}) and minor limit penalties (∑penalties\sum\text{penalties}) are applied broadly across tasks and abbreviated where standard.
Task Term Formula Weight
1. Stand Pose preservation exp⁡(−‖qjoints−q0,joints‖)\exp(-\|q_{\text{joints}}-q_{0,\text{joints}}\|) 1010
Absolute stillness exp⁡(−10.0​‖𝐯CoM‖)⋅exp⁡(−0.1​‖q˙‖)\exp(-10.0\|\mathbf{v}_{\text{CoM}}\|)\cdot\exp(-0.1\|\dot{q}\|) 1515
Feet planted exp⁡(−10.0​(‖𝐯foot,L‖+‖𝐯foot,R‖))\exp(-10.0(\|\mathbf{v}_{\text{foot,L}}\|+\|\mathbf{v}_{\text{foot,R}}\|)) 1010
Torso balance exp⁡(−5.0​max⁡(0,1−𝐮torso⋅𝐳^))\exp(-5.0\max(0,1-\mathbf{u}_{\text{torso}}\cdot\mathbf{\hat{z}})) 1010
Survival 𝕀alive\mathbb{I}_{\text{alive}} 55
2. Powerlift Hold stability 𝕀⁡(zbar>1.6​m∧|z˙bar|<0.1​m/s)\mathbb{I}(z_{\text{bar}}>1.6\text{m}\land|\dot{z}_{\text{bar}}|<0.1\text{m/s}) 100100
Height threshold 𝕀⁡(zbar>1.6​m)\mathbb{I}(z_{\text{bar}}>1.6\text{m}) 5050
Upright alignment exp⁡(−20.0​(1−𝐮torso⋅𝐳^))\exp(-20.0(1-\mathbf{u}_{\text{torso}}\cdot\mathbf{\hat{z}})) 100100
Lumbar spine safety exp⁡(−15.0​|qlumbar−(−0.4)|)\exp(-15.0|q_{\text{lumbar}}-(-0.4)|) 8080
Muscular effort mean​(a)\text{mean}(a) 3030
Hip hinge posture 𝕀⁡(qlumbar>0.1)\mathbb{I}(q_{\text{lumbar}}>0.1) 200200
3. Singleleg Balance (CoM) exp⁡(−15.0​‖𝐩pelvisx​y−𝐩foot,Rx​y‖2)\exp(-15.0\|\mathbf{p}_{\text{pelvis}}^{xy}-\mathbf{p}_{\text{foot,R}}^{xy}\|^{2}) 500500
Lifted foot clear clip​(zfoot,L/0.3,0,1)\text{clip}(z_{\text{foot,L}}/0.3,0,1) 300300
Absolute stillness exp⁡(−5.0​‖𝐯CoM‖2)\exp(-5.0\|\mathbf{v}_{\text{CoM}}\|^{2}) 400400
Upright alignment exp⁡(−10.0​(1−𝐮torso⋅𝐳^)2)\exp(-10.0(1-\mathbf{u}_{\text{torso}}\cdot\mathbf{\hat{z}})^{2}) 200200
Foot touch penalty 𝕀⁡(zfoot,L<0.05​m)\mathbb{I}(z_{\text{foot,L}}<0.05\text{m}) −800-800
Lateral lean penalty |𝐮torso⋅𝐲^||\mathbf{u}_{\text{torso}}\cdot\mathbf{\hat{y}}| −200-200
4. Sit Chair approach 2.0/(‖Δ​𝐩chairx​y‖+0.2)2.0/(\|\Delta\mathbf{p}_{\text{chair}}^{xy}\|+0.2) 1010
3D Contact docking exp⁡(−Δ​x22⋅0.152)⋅exp⁡(−Δ​z22⋅0.152)\exp(-\frac{\Delta x^{2}}{2\cdot 0.15^{2}})\cdot\exp(-\frac{\Delta z^{2}}{2\cdot 0.15^{2}}) 3030
Lateral centering exp⁡(−Δ​y22⋅0.052)\exp(-\frac{\Delta y^{2}}{2\cdot 0.05^{2}}) 3030
Seated leg geometry exp⁡(−2.0​|q¯hip−1.5|−kknee​|q¯knee−1.6|)\exp(-2.0|\bar{q}_{\text{hip}}-1.5|-k_{\text{knee}}|\bar{q}_{\text{knee}}-1.6|) 4040
Feet tucked under exp(−5.0∑i|(xfoot,i−xpelvis)−0.2|)\exp(-5.0\sum_{i}|(x_{\text{foot},i}-x_{\text{pelvis}})-0.2|) 3030
Soft landing exp⁡(−2.0​|z˙pelvis|)\exp(-2.0|\dot{z}_{\text{pelvis}}|) 55
5. Balance Forward lean stance exp⁡(−10.0​(θpitch−(−0.15))2)\exp(-10.0(\theta_{\text{pitch}}-(-0.15))^{2}) (gated) 1515
CoM over board exp⁡(−20.0​‖𝐩CoMx​y−𝐩boardx​y‖2)\exp(-20.0\|\mathbf{p}_{\text{CoM}}^{xy}-\mathbf{p}_{\text{board}}^{xy}\|^{2}) 1515
Board over ball exp⁡(−5.0​‖𝐩boardx​y−𝐩ballx​y‖)\exp(-5.0\|\mathbf{p}_{\text{board}}^{xy}-\mathbf{p}_{\text{ball}}^{xy}\|) 55
Board level exp⁡(−5.0​(1−𝐮board⋅𝐳^))\exp(-5.0(1-\mathbf{u}_{\text{board}}\cdot\mathbf{\hat{z}})) 55
Stand tall exp⁡(−10.0​(1.30−zpelvis)2)\exp(-10.0(1.30-z_{\text{pelvis}})^{2}) if z<1.3z<1.3 44
6. Squat Height tracking exp⁡(−30.0​(zpelvis−ztarget)2)\exp(-30.0(z_{\text{pelvis}}-z_{\text{target}})^{2}) 200200
Joint tracking exp⁡(−5.0​(q¯−qtarget)2)\exp(-5.0(\bar{q}-q_{\text{target}})^{2}) 300/200300/200
Vertical vel tracking exp⁡(−5.0​(vz,CoM−vz,target)2)\exp(-5.0(v_{z,\text{CoM}}-v_{z,\text{target}})^{2}) 200200
CoM centered exp⁡(−15.0​‖𝐩pelvisx​y−𝐩mid_feetx​y‖2)\exp(-15.0\|\mathbf{p}_{\text{pelvis}}^{xy}-\mathbf{p}_{\text{mid\_feet}}^{xy}\|^{2}) 300300
Kneeling penalty 𝕀⁡(min⁡(zknee,L,zknee,R)<0.25​m)\mathbb{I}(\min(z_{\text{knee,L}},z_{\text{knee,R}})<0.25\text{m}) −1000-1000
TABLE VII: Reward Functions for Musculoskeletal Benchmark Tasks (Part 2: Locomotion Series).
Task Term Formula Weight
1. Walk Velocity tracking exp⁡(−2.0​‖𝐯CoM−𝐯target‖2)\exp(-2.0\|\mathbf{v}_{\text{CoM}}-\mathbf{v}_{\text{target}}\|^{2}) 1010
Kinematic imitation exp⁡(−2.0​‖qjoints−qref,joints‖2)\exp(-2.0\|q_{\text{joints}}-q_{\text{ref,joints}}\|^{2}) 33
Upright / No hunch exp⁡(−10.0​(1−𝐮torso⋅𝐳^))\exp(-10.0(1-\mathbf{u}_{\text{torso}}\cdot\mathbf{\hat{z}})), exp⁡(−20.0​(𝐮chest⋅𝐳^+0.1)2)\exp(-20.0(\mathbf{u}_{\text{chest}}\cdot\mathbf{\hat{z}}+0.1)^{2}) 6/86/8
Forward face exp⁡(−5.0​(1−𝐮chest⋅𝐱^))\exp(-5.0(1-\mathbf{u}_{\text{chest}}\cdot\mathbf{\hat{x}})) 88
2. Crawl Backwards velocity clip(−vx,CoM/vtarget,−0.2,1.0)\text{clip}(-v_{x,\text{CoM}}/v_{\text{target}},-0.2,1.0) 500500
Prone posture exp⁡(−5​(𝐮chest⋅𝐳^+1)2)⋅…\exp(-5(\mathbf{u}_{\text{chest}}\cdot\mathbf{\hat{z}}+1)^{2})\cdot\dots 100100
Alternating stride clip​(|Δ​xhands|/0.3,0,1)⋅…\text{clip}(|\Delta x_{\text{hands}}|/0.3,0,1)\cdot\dots (diag gait) 150150
Drift penalty 2.0​|vy,CoM|+|vx,CoM|​𝕀​(vx>0)2.0|v_{y,\text{CoM}}|+|v_{x,\text{CoM}}|\mathbb{I}(v_{x}>0) −150-150
3. Run High velocity clip​(vx,CoM,−0.5,6.0)\text{clip}(v_{x,\text{CoM}},-0.5,6.0), bonus if v>3.0v>3.0 200/100200/100
Asymmetric stride clip​(dx,feet+2.0⋅dz,feet,0.0,1.0)\text{clip}(d_{x,\text{feet}}+2.0\cdot d_{z,\text{feet}},0.0,1.0) 200200
Knee flexion (swing) clip​((max⁡(qknee)−0.1)/1.0,0,1)\text{clip}((\max(q_{\text{knee}})-0.1)/1.0,0,1) 200200
Scissor penalty clip​((0.15−dy,feet)/0.15,0,1)\text{clip}((0.15-d_{y,\text{feet}})/0.15,0,1) −300-300
Kangaroo penalty Penalize dx,feet<0.2∧dz,feet<0.1d_{x,\text{feet}}<0.2\land d_{z,\text{feet}}<0.1 −300-300
4. Jump Flight gradient clip​(12​(zfoot,L+zfoot,R)/0.1,0,1)\text{clip}(\frac{1}{2}(z_{\text{foot,L}}+z_{\text{foot,R}})/0.1,0,1) 600600
Upward explosion clip​(z˙pelvis,0,2.0)\text{clip}(\dot{z}_{\text{pelvis}},0,2.0) 300300
Feet separation pen. dx​y,feetd_{xy,\text{feet}} −200-200
Desync jump pen. |qknee,L−qknee,R|+|qhip,L−qhip,R||q_{\text{knee,L}}-q_{\text{knee,R}}|+|q_{\text{hip,L}}-q_{\text{hip,R}}| −150-150
5. Walk turn Targeted velocity clip​(𝐯CoM⋅𝐮target/3.0,−1.0,1.0)\text{clip}(\mathbf{v}_{\text{CoM}}\cdot\mathbf{u}_{\text{target}}/3.0,-1.0,1.0) 200200
Waypoint approach exp⁡(−‖Δ​𝐩target‖)\exp(-\|\Delta\mathbf{p}_{\text{target}}\|) 5050
Target facing clip​(𝐮fwd⋅𝐮target,0,1)\text{clip}(\mathbf{u}_{\text{fwd}}\cdot\mathbf{u}_{\text{target}},0,1) 5050
Waypoint reached 𝕀⁡(‖Δ​𝐩target‖<0.5​m)\mathbb{I}(\|\Delta\mathbf{p}_{\text{target}}\|<0.5\text{m}) 500500
Wall collision 𝕀contact​(wall)\mathbb{I}_{\text{contact}}(\text{wall}) −200-200
6. Sidestep Lateral velocity clip​(vy,CoM/vtarget,−0.2,1.0)\text{clip}(v_{y,\text{CoM}}/v_{\text{target}},-0.2,1.0) 400400
Phase sync exp⁡(−15.0​(eL+eR))\exp(-15.0(e_{L}+e_{R})) 200200
CoM mid-feet lock |ypelvis−12​(yfoot,L+yfoot,R)||y_{\text{pelvis}}-\frac{1}{2}(y_{\text{foot,L}}+y_{\text{foot,R}})| −500-500
Rubber band splits max⁡(0,dx​y,feet−0.3)\max(0,d_{xy,\text{feet}}-0.3) −400-400
TABLE VIII: Reward Functions for Musculoskeletal Benchmark Tasks (Part 3: Interaction with Environment Series).
Task Term Formula Weight
1. Stairs Fwd trajectory clip​(vx,CoM,−0.2,2.0)\text{clip}(v_{x,\text{CoM}},-0.2,2.0) 5050
Upward trajectory clip​(vz,CoM,−0.5,3.0)\text{clip}(v_{z,\text{CoM}},-0.5,3.0) 3030
Vertical stable exp⁡(−5.0​(0.98−𝐮torso⋅𝐳^)2)\exp(-5.0(0.98-\mathbf{u}_{\text{torso}}\cdot\mathbf{\hat{z}})^{2}) 2020
Lateral stable exp⁡(−5.0​ypelvis2)\exp(-5.0y_{\text{pelvis}}^{2}) 2020
Anticipatory lift 𝕀⁡(radar)⋅max⁡(z˙swing,0)\mathbb{I}(\text{radar})\cdot\max(\dot{z}_{\text{swing}},0) 4040
2. Hurdle Forward velocity clip​(vx,CoM,−0.1,4.0)\text{clip}(v_{x,\text{CoM}},-0.1,4.0) 300300
Clearance (Flight) 𝕀⁡(max⁡(zfeet)>htarget)\mathbb{I}(\max(z_{\text{feet}})>h_{\text{target}}) 100100
Sparse clear bonus 𝕀⁡(Δ​x<0∧|ypelvis|<0.5​m)\mathbb{I}(\Delta x<0\land|y_{\text{pelvis}}|<0.5\text{m}) 30003000
Crane hopping pen. clip​((|zfoot,L−zfoot,R|−0.2)/0.5,0,1)\text{clip}((|z_{\text{foot,L}}-z_{\text{foot,R}}|-0.2)/0.5,0,1) −150-150
3. Step stones Target stone progress Bonus when crossing stone X-coord 800800
Magnetic reaching exp(−2.0⋅dlead)⋅𝕀(vx,CoM>0.1)\exp(-2.0\cdot d_{\text{lead}})\cdot\mathbb{I}(v_{x,\text{CoM}}>0.1) 150150
Stand tall clip​((Δ​zleg−0.6)/0.3,0,1)\text{clip}((\Delta z_{\text{leg}}-0.6)/0.3,0,1) 5050
Overbend knee pen. max⁡(qknee−1.0,0)\max(q_{\text{knee}}-1.0,0) −50-50
4. Slide Pitch balance Rgate⋅exp(−5.0⋅θpitch2)R_{\text{gate}}\cdot\exp(-5.0\cdot\theta_{\text{pitch}}^{2}) 22
Stand tall Rgate⋅exp⁡(−10.0​(0.95−Δ​zleg)2)R_{\text{gate}}\cdot\exp(-10.0(0.95-\Delta z_{\text{leg}})^{2}) 22
Slope forward lean clip(−𝐮chest⋅𝐳^/0.3,0,1)\text{clip}(-\mathbf{u}_{\text{chest}}\cdot\mathbf{\hat{z}}/0.3,0,1) 2020
Backward lean pen. max⁡(𝐮chest⋅𝐳^,0)\max(\mathbf{u}_{\text{chest}}\cdot\mathbf{\hat{z}},0) −100-100
5. Open door Hand attractor field exp⁡(−1.5​‖Δ​𝐩handle‖)\exp(-1.5\|\Delta\mathbf{p}_{\text{handle}}\|) 400400
Open hinge clip​(θdoor/1.2,0,1)⋅prox_gate\text{clip}(\theta_{\text{door}}/1.2,0,1)\cdot\text{prox\_gate} 500500
Pass threshold Sparse bonuses for x>1.0,2.0x>1.0,2.0 20002000
Dive/Split penalties Penalize pelvis lead or wide feet −300/−400-300/-400
6. Reach Dense reach dist exp⁡(−4.0​‖Δ​𝐩reach‖)\exp(-4.0\|\Delta\mathbf{p}_{\text{reach}}\|) 5050
Target velocity proj clip​(𝐯hand⋅𝐮reach,0,2.0)\text{clip}(\mathbf{v}_{\text{hand}}\cdot\mathbf{u}_{\text{reach}},0,2.0) 2020
Target touch bonus 𝕀⁡(‖Δ​𝐩reach‖<0.15​m)\mathbb{I}(\|\Delta\mathbf{p}_{\text{reach}}\|<0.15\text{m}) 200200
7. Walk & sit Backwards precision exp(−∥Δ𝐩chairx​y∥2/0.005)\exp(-\|\Delta\mathbf{p}_{\text{chair}}^{xy}\|^{2}/0.005) 100100
Fast descent bonus 𝕀⁡(z˙pelvis<−1.5​m/s)\mathbb{I}(\dot{z}_{\text{pelvis}}<-1.5\text{m/s}) 5050
Relaxed recline 𝕀⁡(𝐮torso⋅𝐳^<0.8)\mathbb{I}(\mathbf{u}_{\text{torso}}\cdot\mathbf{\hat{z}}<0.8) (once seated) 200200
90-degree pose exp⁡(−5​(q¯hip−1.5)2)+…\exp(-5(\bar{q}_{\text{hip}}-1.5)^{2})+\dots 100100
8. Chin up Spatial gradient exp⁡(−5.0​max⁡(0,hbar−zhead))\exp(-5.0\max(0,h_{\text{bar}}-z_{\text{head}})) 5050
Muscle force filter mean​(Fpull_muscles)\text{mean}(F_{\text{pull\_muscles}}) 55
Explosive accel max⁡(0,z¨pelvis)\max(0,\ddot{z}_{\text{pelvis}}) 4040
Anti-kipping wiggle ‖q˙legs‖\|\dot{q}_{\text{legs}}\| −20-20
9. Catch Hand-object reach exp(−3.0⋅dcatch)\exp(-3.0\cdot d_{\text{catch}}) 100100
Interception bonus 𝕀⁡(dcatch<0.15​m∧zcube>0.5​m)\mathbb{I}(d_{\text{catch}}<0.15\text{m}\land z_{\text{cube}}>0.5\text{m}) 500500
Drop penalty 𝕀⁡(zcube<0.1​m)\mathbb{I}(z_{\text{cube}}<0.1\text{m}) −200-200
10. Pole walk Target feet phases exp⁡(−10.0​(Δ​xfoot±0.25​sin⁡ϕ)2)\exp(-10.0(\Delta x_{\text{foot}}\pm 0.25\sin\phi)^{2}) 400400
Dynamic radar gap E.g., clip​(2.0⋅vy,CoM,0,1)\text{clip}(2.0\cdot v_{y,\text{CoM}},0,1) if left gap 800800
Fatal collision 𝕀fatal_contact\mathbb{I}_{\text{fatal\_contact}} (head/torso to pole) −1500-1500
TABLE IX: Task Settings, Specific Observations, and Termination Criteria (Part 1: Stabilization Series). Base Observation (ObaseO_{\text{base}}): All tasks fundamentally include proprioceptive states (q,q˙,𝐯CoM)(q,\dot{q},\mathbf{v}_{\text{CoM}}) and muscle states (length, velocity, force, activation). Only task-specific exteroceptive or goal variables are listed below.
Task Specific Obs (OtaskO_{\text{task}}) Initialization (Brief) Termination Criteria
1. Stand Torso quat, feet height Default neutral, zero vel. zCoM<0.8z_{\text{CoM}}<0.8 or tilt >0.8>0.8
2. Powerlift Dumbbell 𝐩bar\mathbf{p}_{\text{bar}}, Δ​𝐩bar\Delta\mathbf{p}_{\text{bar}}, tilt High CoM, pre-loaded lumbar zpelvis<0.7z_{\text{pelvis}}<0.7 or qlumbar>0.5q_{\text{lumbar}}>0.5
3. Singleleg Torso quat, feet/CoM zz Flamingo pose, noise zpelvis<0.6z_{\text{pelvis}}<0.6, dropped foot
4. Sit Chair relative Δ​𝐩chair\Delta\mathbf{p}_{\text{chair}} In front of chair, hips pre-flexed zpelvis<0.25z_{\text{pelvis}}<0.25, invalid contact
5. Balance Board 𝐩,𝐪\mathbf{p},\mathbf{q}, CoM/Ball Δ​𝐩\Delta\mathbf{p} Drop z=1.28z=1.28m, guard stance zpelvis<0.70z_{\text{pelvis}}<0.70, pitch fall
6. Squat Phase ϕ\phi, target ztargetz_{\text{target}} Squat start pose, barbell hold zpelvis<0.35z_{\text{pelvis}}<0.35, kneeling
TABLE X: Task Settings, Specific Observations, and Termination Criteria (Part 2: Locomotion Series).
Task Specific Obs (OtaskO_{\text{task}}) Initialization (Brief) Termination Criteria
1. Walk Phase ϕ\phi, error Δ​q=q−qref\Delta q=q-q_{\text{ref}} Sampled from synthetic walk zCoM<0.5z_{\text{CoM}}<0.5 or tilt >0.8>0.8
2. Crawl Torso quat, feet height Prone quadruped, ground z=0.25z=0.25 Fwd move, twist, flip
3. Run Torso angle, feet height Standard standing zpelvis<0.45z_{\text{pelvis}}<0.45, fall backward
4. Jump Torso angle, dx​y,feetd_{xy,\text{feet}} Micro-squat, pre-flexed zpelvis<0.45z_{\text{pelvis}}<0.45, wide split
5. Walk turn Target Δ​𝐩\Delta\mathbf{p}, ID, facing aa Stand, 20-step novice protection zpelvis<0.55z_{\text{pelvis}}<0.55, zhead<0.8z_{\text{head}}<0.8
6. Sidestep Phase ϕ\phi, lat dist Δ​yfeet\Delta y_{\text{feet}} Crab guard stance, hips wide zpelvis<0.45z_{\text{pelvis}}<0.45, split >0.5>0.5m
TABLE XI: Task Settings, Specific Observations, and Termination Criteria (Part 3: Interaction with Environment Series).
Task Specific Obs (OtaskO_{\text{task}}) Initialization (Brief) Termination Criteria
1. Stairs Phase ϕ\phi, radar 𝐳terrain∈ℝ9\mathbf{z}_{\text{terrain}}\in\mathbb{R}^{9} Walk posture, initial fwd vel zpelvis<0.75z_{\text{pelvis}}<0.75, lat drift
2. Hurdle Phase ϕ\phi, nearest Δ​𝐡\Delta\mathbf{h} Standing, scan dynamic hurdles zpelvis<0.55z_{\text{pelvis}}<0.55, heavy yaw
3. Step stones Phase ϕ\phi, stones Δ​𝐩stones\Delta\mathbf{p}_{\text{stones}} Elevated platform z=3.05z=3.05m Fall from height, zpelvis<2.5z_{\text{pelvis}}<2.5
4. Slide Phase ϕ\phi, normal 𝐧slope\mathbf{n}_{\text{slope}} Elevated over wave terrain Backwards fall, split >1.2>1.2m
5. Open door Handle Δ​𝐩\Delta\mathbf{p}, door θ\theta, target Stand front of closed door zpelvis<0.65z_{\text{pelvis}}<0.65, head <1.1<1.1
6. Reach Phase ϕ\phi, reach Δ​𝐩reach\Delta\mathbf{p}_{\text{reach}} Quarter-squat, hands raised Knee drop <0.15<0.15m
7. Walk & sit Chair Δ​𝐩\Delta\mathbf{p}, phase 𝕀seated\mathbb{I}_{\text{seated}} Face away, inject backward vel Ground contact (head/knees)
8. Chin up Phase (0), bar dist Δ​zchin\Delta z_{\text{chin}} Hanging, flexed shoulders Fall off bar (zpelvis<0.5z_{\text{pelvis}}<0.5)
9. Catch Cube 𝐩,𝐯\mathbf{p},\mathbf{v}, dist dcatchd_{\text{catch}} Anticipatory basket, cube spawn Drop cube (zcube<0.05z_{\text{cube}}<0.05)
10. Pole walk Phase ϕ\phi, radar 𝐳radar∈ℝ7\mathbf{z}_{\text{radar}}\in\mathbb{R}^{7} Elevated 1.051.05m, high core gain Fatal upper-body collision

A-B Evaluation Details

Full-suite success and robustness use Table XII’s common code-level criteria for the four reward-based baselines. The focused MuscleMimic–DepRL study has its own matched protocol and is excluded from the aggregate robustness scores. Task rewards are never pooled across tasks. Tables XIII–XV expose per-task success, perturbation response, activation and smoothness, including cases where a good aggregate hides a failure.

A-B 1 Success Rate

Success is evaluated separately from environment termination. A positive native solved signal counts; otherwise the evaluator checks the task-specific condition over the recorded history. Survival-horizon tasks explicitly require reaching the horizon without early termination. For other tasks, early termination prevents further progress but is not an additional universal veto. The reported family and overall rates average tasks equally; coverage counts tasks with observed success above zero.

Equation 3
Success Rate\displaystyle\text{Success Rate} =1N​∑i=1N𝕀⁡(𝒮task(i)),\displaystyle=\frac{1}{N}\sum_{i=1}^{N}\mathbb{I}\!\left(\mathcal{S}_{\mathrm{task}}^{(i)}\right), (3)
TABLE XII: Task-specific success criteria used by the shared full-suite evaluator for PPO, SAC, DepRL, and DynSyn-SAC. These criteria are distinct from the environment termination rules. A positive native solved signal also counts as success for every task.
Task Success criterion Task Success criterion
Stand Reach 1,000 evaluation steps without earlier termination. Sidestep Maximum absolute lateral pelvis displacement from the start is at least 9.09.0 m.
Powerlift Reach 900 steps and attain a maximum dumbbell height of at least 1.61.6 m. Stairs Maximum pelvis height is at least 1.61.6 m.
Single-leg stand Reach 1,000 evaluation steps without earlier termination. Hurdle The recorded number of cleared hurdles reaches the environment’s total (five if unavailable).
Sit The environment’s has_sat_down flag becomes true. Step stones current_stone_idx reaches the final stone index.
Balance Run for more than 350 steps and attain a board–ball horizontal separation below 0.50.5 m. Slide Maximum forward pelvis displacement from the start is at least 18.018.0 m.
Squat Pelvis height first falls below 0.750.75 m and subsequently rises above 0.900.90 m. Door open Forward displacement is at least 2.02.0 m, the door-open flag is true, or maximum door angle exceeds 1.01.0 rad.
Walk Reach 1,000 evaluation steps without earlier termination. Reach Accumulate at least two target touches.
Crawl Maximum backward pelvis displacement from the start is at least 2.02.0 m. Walk and sit The sit flag becomes true and the rollout reaches 500 steps.
Run Maximum forward pelvis displacement from the start is at least 15.015.0 m. Chin-up Keep the head more than 0.10.1 m above the bar for at least 25 consecutive steps.
Jump Complete at least 10 hops; a hop is counted after pelvis height crosses above 1.001.00 m and then below 0.950.95 m. Catch The catch flag becomes true and the rollout reaches 200 steps.
Walk turn current_target_idx reaches at least 2. Pole walk Maximum forward pelvis displacement from the start is at least 1.61.6 m.

A-B 2 Robustness

Action and observation tests add Gaussian noise, then clip to their valid ranges. Dynamics tests independently scale each muscle’s force gain by a uniform factor in [1−σ, 1+σ], fixed throughout an episode. Action/dynamics scales are 0, 0.05, 0.10, 0.15 and 0.20; observation scales are 0, 0.02, 0.05, 0.08 and 0.10. Robustness is trapezoidal area under success versus scale, normalized by the tested range and averaged equally across tasks and perturbation types. These separate sweeps characterize tested disturbances, not unrestricted robustness.

Equations 4–6
𝐮t′\displaystyle\mathbf{u}^{\prime}_{t} =clip[−1,1]⁡(𝐮t+ϵtact),\displaystyle=\operatorname{clip}_{[-1,1]}\left(\mathbf{u}_{t}+\bm{\epsilon}_{t}^{\mathrm{act}}\right), (4)
ϵtact\displaystyle\bm{\epsilon}_{t}^{\mathrm{act}} ∼𝒩⁡(𝟎,σact2​𝐈),\displaystyle\sim\mathcal{N}\left(\mathbf{0},\sigma_{\mathrm{act}}^{2}\mathbf{I}\right),
𝐬~t′\displaystyle\tilde{\mathbf{s}}^{\prime}_{t} =clip[−20,20]⁡(𝐬~t+ϵtobs),\displaystyle=\operatorname{clip}_{[-20,20]}\left(\tilde{\mathbf{s}}_{t}+\bm{\epsilon}_{t}^{\mathrm{obs}}\right), (5)
ϵtobs\displaystyle\bm{\epsilon}_{t}^{\mathrm{obs}} ∼𝒩⁡(𝟎,σobs2​𝐈),\displaystyle\sim\mathcal{N}\left(\mathbf{0},\sigma_{\mathrm{obs}}^{2}\mathbf{I}\right),
δm\displaystyle\delta_{m} ∼𝒰⁡(−σdyn,σdyn),\displaystyle\sim\mathcal{U}\left(-\sigma_{\mathrm{dyn}},\sigma_{\mathrm{dyn}}\right), (6)
rm\displaystyle r_{m} =1+δm,\displaystyle=1+\delta_{m},
Gainm′\displaystyle\mathrm{Gain}^{\prime}_{m} =Gainm×rm,\displaystyle=\mathrm{Gain}_{m}\times r_{m},

A-B 3 Activation Cost

Activation cost averages squared actuator activation over muscles and time. Lower values indicate less activation effort only for successful, behaviorally comparable policies. This proxy does not directly measure metabolic energy, fatigue or physiological fidelity.

Equation 7
Emuscle\displaystyle E_{\mathrm{muscle}} =1T​∑t=1T(1M​∑m=1Mam,t2),\displaystyle=\frac{1}{T}\sum_{t=1}^{T}\left(\frac{1}{M}\sum_{m=1}^{M}a_{m,t}^{2}\right), (7)

A-B 4 Joint Smoothness

The implementation takes a second finite difference of joint angular velocity divided by the squared timestep, yielding angular jerk. It then averages the squared quantity over joints/time and reports its base-ten logarithm. The pre-log units are rad²/s⁶. Interpret it within successful comparable behaviors; a failed or motionless controller can appear smooth without solving the task.

Equations 8–10
D(2)​vj,t\displaystyle\mathrm{D}^{(2)}v_{j,t} =vj,t−2​vj,t−1+vj,t−2Δ​t2\displaystyle=\frac{v_{j,t}-2v_{j,t-1}+v_{j,t-2}}{\Delta t^{2}} (8)
≈d2​vj​(t)d​t2=d3​θj​(t)d​t3.\displaystyle\approx\frac{\mathrm{d}^{2}v_{j}(t)}{\mathrm{d}t^{2}}=\frac{\mathrm{d}^{3}\theta_{j}(t)}{\mathrm{d}t^{3}}.
Jsq\displaystyle J_{\mathrm{sq}} =1(N−2)​Njoint​∑t=3N∑j=1Njoint(D(2)​vj,t)2,\displaystyle=\frac{1}{(N-2)N_{\mathrm{joint}}}\sum_{t=3}^{N}\sum_{j=1}^{N_{\mathrm{joint}}}\left(\mathrm{D}^{(2)}v_{j,t}\right)^{2}, (9)
Smoothnessjerk\displaystyle\mathrm{Smoothness}_{\mathrm{jerk}} =log10⁡(Jsq+ϵ),\displaystyle=\log_{10}\left(J_{\mathrm{sq}}+\epsilon\right), (10)

A-B 5 Peak Efficiency Steps

Peak efficiency steps are the evaluated training step attaining maximum mean return over the complete training horizon, paired with that return. This retrospective argmax is not a first-success threshold or a general sample-complexity guarantee. Different horizons, checkpoint schedules and task reward scales affect its interpretation.

Equations 11–12
Tmax\displaystyle T_{\mathrm{max}} =arg​maxt∈𝒯⁡𝔼​[Rt],\displaystyle=\operatorname*{arg\,max}_{t\in\mathcal{T}}\mathbb{E}\left[R_{t}\right], (11)
Rmax\displaystyle R_{\mathrm{max}} =maxt∈𝒯⁡𝔼⁡[Rt].\displaystyle=\max_{t\in\mathcal{T}}\mathbb{E}\left[R_{t}\right]. (12)

A-B 6 EMG Similarity

Walking and stairs use Gait120; running uses 3-m/s recordings from Van Hooren and Meijer [Gait120], [running dataset]. Table V reports three selected walking muscles individually, eleven running channels averaged and twelve stair muscles averaged. Profiles are resampled, normalized and cyclically shifted to maximize correlation. The score measures phase-optimized envelope shape, not absolute activation amplitude or event timing.

Equations 13–14
rm\displaystyle r_{m} =∑t=1T(asim,m​(t)−a¯sim,m)​(eEMG,m​(t)−e¯EMG,m)∑t=1T(asim,m​(t)−a¯sim,m)2\displaystyle=\frac{\sum_{t=1}^{T}\left(a_{\mathrm{sim},m}(t)-\bar{a}_{\mathrm{sim},m}\right)\left(e_{\mathrm{EMG},m}(t)-\bar{e}_{\mathrm{EMG},m}\right)}{\sqrt{\sum_{t=1}^{T}\left(a_{\mathrm{sim},m}(t)-\bar{a}_{\mathrm{sim},m}\right)^{2}}} (13)
×1∑t=1T(eEMG,m​(t)−e¯EMG,m)2,\displaystyle\times\frac{1}{\sqrt{\sum_{t=1}^{T}\left(e_{\mathrm{EMG},m}(t)-\bar{e}_{\mathrm{EMG},m}\right)^{2}}},
rEMG\displaystyle r_{\mathrm{EMG}} =1M​∑m=1Mrm,\displaystyle=\frac{1}{M}\sum_{m=1}^{M}r_{m}, (14)
TABLE XIII: Benchmark Performance and Robustness Evaluation: Stabilization. Train-env. SR retains native training noise; robustness columns report separate perturbation sweeps (App. A-B).
Task Algorithm Activation Robustness (%) Max Train-env. log10\log_{10}smooth
Cost Action Observation Dynamic Steps (×107\times 10^{7}) SR (%) (rad2/s6)
0 0.05 0.1 0.15 0.2 0 0.02 0.05 0.08 0.10 0 0.05 0.1 0.15 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
TABLE XIV: Benchmark Performance and Robustness Evaluation: Locomotion. Train-env. SR retains native training noise; robustness columns report separate perturbation sweeps (App. A-B).
Task Algorithm Activation Robustness(%) Max Train-env. log10\log_{10}smooth
Cost Action Observation Dynamic Steps(×107\times 10^{7}) SR (%) (rad2/s6)
0 0.05 0.1 0.15 0.2 0 0.02 0.05 0.08 0.10 0 0.05 0.1 0.15 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
TABLE XV: Benchmark Performance and Robustness Evaluation: Interaction with Environment. Train-env. SR retains native training noise; robustness columns report separate perturbation sweeps (App. A-B).
Task Algorithm Activation Robustness (%) Max Train-env. log10\log_{10}smooth
Cost Action Observation Dynamic Steps (×107\times 10^{7}) SR (%) (rad2/s6)
0 0.05 0.1 0.15 0.2 0 0.02 0.05 0.08 0.10 0 0.05 0.1 0.15 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

A-C Detailed Descriptions of Reward-Based RL Baselines

PPO uses clipped on-policy updates; SAC uses off-policy maximum-entropy learning. DepRL combines the policy with Differential Extrinsic Plasticity exploration driven by temporal sensorimotor relationships. DynSyn-SAC learns group-level synergistic actions plus state-dependent muscle-level corrections, with a schedule controlling their strength. The reported benefits depend on these embodiments and training settings; the descriptions do not themselves establish convergence or physiological correctness.

A-C 1 Training and Checkpoint Protocol

Table XVI records released defaults, all with seed zero. PPO/SAC use eight parallel training environments, DepRL sixteen and DynSyn-SAC sixty-four. PPO/SAC budgets are 200 million steps, DepRL 100 million and DynSyn-SAC 50 million; checkpoint rules also differ. Off-policy updates per vectorized collection step must not be confused with updates per scalar transition. The success scripts evaluate the selected checkpoint for fifty episodes per task by default, rather than choosing checkpoints by success. These details limit claims of a fully compute-matched algorithm comparison.

TABLE XVI: Released training and checkpoint protocol for the four reward-based baselines. MLP widths list hidden layers for the actor and value/critic networks. “Det.” denotes deterministic evaluation episodes.
Method Actor / critic MLP Learning rate Batch Collection, rollout, and update/data cadence Budget Checkpoint used for reporting
PPO 2×10242\times 1024 / 2×10242\times 1024 3×10−43\times 10^{-4} 256 2,048 steps per environment and eight environments (16,38416{,}384 transitions per rollout); 10 epochs, i.e., 10 passes over each rollout. 2×1082\times 10^{8} steps Highest mean evaluation return; 10 det. episodes every 5×1055\times 10^{5} training steps.
SAC 2×10242\times 1024 / twin 2×10242\times 1024 Q networks 1×10−41\times 10^{-4} 256 Continuous off-policy collection; one vectorized step (eight new transitions) followed by one gradient step. The configured gradient_steps/train_freq ratio is 1. 2×1082\times 10^{8} steps Highest mean evaluation return; 10 det. episodes every 2×1052\times 10^{5} training steps.
DepRL 2×10242\times 1024 / 2×10242\times 1024 Actor 5×10−55\times 10^{-5}; critic 8×10−58\times 10^{-5}; dual 2×10−32\times 10^{-3} 256 Continuous replay with 3-step returns; after 10510^{5} warm-up transitions, 30 minibatch updates per 1,000 new transitions (0.03 update iterations per transition). 1×1081\times 10^{8} steps Last step-numbered checkpoint (the completed run ends at 10810^{8}); saved every 2×1062\times 10^{6} steps. A best-return checkpoint is also written, but is not the evaluator default.
DynSyn-SAC 3×10243\times 1024 / twin 3×10243\times 1024 Q networks, plus DynSyn layer Linear 5×10−45\times 10^{-4} to 1.25×10−41.25\times 10^{-4} 256* Continuous off-policy collection; one vectorized step (64 new transitions) followed by one gradient step. The configured gradient_steps/train_freq ratio is 1. 5×1075\times 10^{7} steps Highest mean evaluation return; 3 det. episodes every 5×1065\times 10^{6} training steps.

*DynSyn-SAC does not override batch_size; 256 is the Stable-Baselines3 SAC default in the declared 2.x dependency range. It likewise inherits a one-step train_freq; the released configuration explicitly sets one gradient step.

A-D Detailed Muscle Activation Analysis: 416 vs. 700 Muscle Models

The complete 416- and 700-muscle powerlift heatmaps show model-specific recruitment. Additional neck, deep-trunk and scapular actuators in the expanded anatomy alter the available pathways and diagnostic resolution. Their activation patterns do not prove more accurate human force transfer or isolate why recruitment changes.

Refer to caption
Fig. 14: Upper limb muscle activation heatmap for the 416-muscle model (Powerlift). The heatmap displays the intense, sustained activations of the available arm and shoulder musculature required to stabilize the barbell overhead. Note that this model lacks the detailed deep axial stabilizers of the neck and spine.
Refer to caption
Fig. 15: Expanded muscle activation heatmap for the 700-muscle model (Powerlift). The top section includes the upper limb muscles, where the distinct ”blank gaps” represent the fine-grained scapular and arm muscles exclusive to the 700-model. The bottom section explicitly details the heavy recruitment of the newly available neck and spine axial stabilizers (e.g., multifidus, splenius).

A-E Agentic Reward Tuning

An external reasoner periodically receives recent performance statistics and returns reward weights. The original setup uses forward speed and torso pitch every 20,000 steps, with relative coefficient changes bounded by 20% but no global normalization. A shared file synchronizes workers; locking avoids duplicate calls and exception handling lets learning continue after a failed request. The matched DeepSeek/GPT comparison instead fixes the L1 norm at 40 under the same learner and budget. Its metric trade-offs should not be merged with the original unnormalized study.

Equations 15–17
Rt=∑iwi,t​ri,t.R_{t}=\sum_{i}w_{i,t}r_{i,t}. (15)
𝐬stats=Aggregate⁡({𝐨τ,𝐫τ}τ=t−Kt),\mathbf{s}_{\mathrm{stats}}=\mathrm{Aggregate}\left(\{\mathbf{o}_{\tau},\mathbf{r}_{\tau}\}_{\tau=t-K}^{t}\right), (16)
𝐰t+1=Reasoner⁡(𝐰t,𝐬stats),\mathbf{w}_{t+1}=\mathrm{Reasoner}\left(\mathbf{w}_{t},\mathbf{s}_{\mathrm{stats}}\right), (17)

A-F Direct GPT-6 Control: Experimental Details

At each paused decision, the controller sees static metadata and a lossless JSON restructuring of the 2,418-D policy observation: root pose, joints/velocities, 354 muscles with five state values each, contact sensors and a 469-D goal/reference block. It cannot inspect privileged diagnostics, rendered frames or reference-controller actions. A response must contain exactly 354 finite numeric values within [−1,1]; it is applied directly without clipping, smoothing or baseline addition.

The runner resets seed zero and deterministically replays committed actions before extending the rollout. Physics runs at 1 ms, native control at 10 ms and each language-model action is held for 0.2 seconds. The environment pauses during reasoning, so wall-clock delay is not simulated inference latency. Fall checks occur at every native step: nonfinite state, root height below 0.6 m or tilt above 70°. Analysis includes the first fallen sample and excludes later states; the original Walk log contains post-fall states that are explicitly discarded.

Tracking error uses sixteen sites and all native samples in the truncated prefix. Divergence requires five consecutive samples above both the time-aligned frozen-prior reference error and its 95th percentile. The reference is used to define this diagnostic, not as another reported baseline. N/A means the sustained threshold was not reached before the analyzed rollout ended.

TABLE XVII: Controller-visible observation supplied at each GPT-6 decision. Dimensions are identical for Walk, Run, and Stairs.
Observation block Dimension
Root pose without global xx–yy 5
Joint positions 82
Root velocity 6
Joint velocities 82
Muscle state (354×354\times length, velocity, force, excitation, activation) 1,770
Foot/toe touch sensors 4
Goal and reference features 469
Total 2,418
Equations 18–20
θ=cos−1⁡(clip⁡(1−2​(x2+y2),−1,1)).\theta=\cos^{-1}\!\left(\operatorname{clip}\left(1-2(x^{2}+y^{2}),-1,1\right)\right). (18)
et=13​N​∑j=1N‖𝐱j,t−𝐱j,tref‖22,N=16.e_{t}=\sqrt{\frac{1}{3N}\sum_{j=1}^{N}\left\|\mathbf{x}_{j,t}-\mathbf{x}^{\mathrm{ref}}_{j,t}\right\|_{2}^{2}},\qquad N=16. (19)
etGPT>e~trefandetGPT>Q0.95ref.e_{t}^{\mathrm{GPT}}>\widetilde{e}_{t}^{\mathrm{ref}}\quad\text{and}\quad e_{t}^{\mathrm{GPT}}>Q_{0.95}^{\mathrm{ref}}. (20)

A-G Latent-Action RL

A teacher encoder–decoder is pretrained by behavior cloning on expert muscle actions and anatomical features. The decoder conditions on latent intent, physical priors, muscle state and muscle–joint moment information. In hard mode, a raw policy action is normalized to [0,1], passed through the teacher and clipped to a radius-δ tube around the reconstructed action before mapping back to [−1,1]. A decaying deviation penalty discourages raw commands outside that tube. Transformer attention captures inter-muscle dependencies. The tube constrains exploration around a learned prior; it is not an independently verified guarantee of physiological plausibility.

Equations 21–28
A^=Decoder⁡(Encoder⁡(A),P,S,M),\hat{A}=\mathrm{Decoder}(\mathrm{Encoder}(A),P,S,M), (21)
ℒBC=1N​∑i‖A^i−Ai+12‖2.\mathcal{L}_{\mathrm{BC}}=\frac{1}{N}\sum_{i}\left\|\hat{A}_{i}-\frac{A_{i}+1}{2}\right\|^{2}. (22)
at,01RL=atRL+12.a_{t,01}^{\mathrm{RL}}=\frac{a_{t}^{\mathrm{RL}}+1}{2}. (23)
a^t=Decoder⁡(Encoder⁡(atRL),Pt,St,Mt),\hat{a}_{t}=\mathrm{Decoder}\left(\mathrm{Encoder}(a_{t}^{\mathrm{RL}}),P_{t},S_{t},M_{t}\right), (24)
at,01env=clip⁡(at,01RL,a^t−δ,a^t+δ),a_{t,01}^{\mathrm{env}}=\mathrm{clip}\left(a_{t,01}^{\mathrm{RL}},\hat{a}_{t}-\delta,\hat{a}_{t}+\delta\right), (25)
atenv=2​at,01env−1.a_{t}^{\mathrm{env}}=2a_{t,01}^{\mathrm{env}}-1. (26)
ℒdev=‖at,01RL−at,01env‖2.\mathcal{L}_{\mathrm{dev}}=\left\|a_{t,01}^{\mathrm{RL}}-a_{t,01}^{\mathrm{env}}\right\|^{2}. (27)
Rt′=Rt−λt​ℒdev,R_{t}^{\prime}=R_{t}-\lambda_{t}\mathcal{L}_{\mathrm{dev}}, (28)

A-H EMG Data Processing and Comparative Analysis

Each human gait cycle is interpolated to 101 samples and min–max normalized. Multiple cycles yield a mean profile and normalized variability. The simulated profile is cyclically shifted to maximize Pearson correlation with the human mean, and the same shift is applied to its variability. This deliberately removes absolute scale and phase; the resulting agreement should be reported as envelope-shape similarity rather than faithful muscle timing or force.

Equations 29–32
x~=interp1d​(xold,xraw,kind=’cubic’)​(xnew),\tilde{x}=\text{interp1d}(x_{\text{old}},x_{\text{raw}},\text{kind='cubic'})(x_{\text{new}}), (29)
xnorm=x−xminxmax−xmin+ϵ,x_{\text{norm}}=\frac{x-x_{\text{min}}}{x_{\text{max}}-x_{\text{min}}+\epsilon}, (30)
σnorm=std​(xcycles)xmax−xmin+ϵ.\sigma_{\text{norm}}=\frac{\text{std}(x_{\text{cycles}})}{x_{\text{max}}-x_{\text{min}}+\epsilon}. (31)
r=cov​(x~sim,x~human)σsim​σhuman.r=\frac{\text{cov}(\tilde{x}_{\text{sim}},\tilde{x}_{\text{human}})}{\sigma_{\text{sim}}\sigma_{\text{human}}}. (32)

A-I Muscle Recruitment Categorization

Average each actuator’s activation over the rollout, then sum those means within name-keyword anatomical groups. Divide each group’s activation mass by the total to obtain the relative share. This is neither a time integral nor an average per actuator, so group size and actuator definitions affect comparisons. Table XVIII preserves the exact keyword grouping, including unclassified muscles. Its categories are implementation rules, not a newly validated anatomical decomposition.

Equations 33–37
a¯m=1T​∑t=1Tam,t,\bar{a}_{m}=\frac{1}{T}\sum_{t=1}^{T}a_{m,t}, (33)
m∈Ck⟺∃q∈𝒬ks.t.q⊆name(m),m\in C_{k}\quad\Longleftrightarrow\quad\exists q\in\mathcal{Q}_{k}\ \mathrm{s.t.}\ q\subseteq\mathrm{name}(m), (34)
Ak=∑m∈Cka¯m.A_{k}=\sum_{m\in C_{k}}\bar{a}_{m}. (35)
Atotal=∑k=1KAk.A_{\mathrm{total}}=\sum_{k=1}^{K}A_{k}. (36)
Pk=AkAtotal×100%.P_{k}=\frac{A_{k}}{A_{\mathrm{total}}}\times 100\%. (37)
TABLE XVIII: Anatomical muscle-family categories used for whole-body recruitment analysis. Each actuator is assigned to a category according to whether its name contains one of the corresponding keywords.
Body Region Muscle Category Actuator Name Keywords
Core and Back Paraspinals (Deep) Ps_
Iliocostalis IL_
Longissimus (Erector Spinae) LTpT, LTpL
Abdominals rect_abd, EO, IO
Other Trunk/Back LAT, LD, QL, MF_m, TRAP, RHOM
Pectorals PEC
Upper Extremity Deltoids DELT
Rotator Cuff SUPSP, INFSP, SUBSC, TMIN, TMAJ, CORB
Elbow Flexors BIC, BRA, BRD
Elbow Extensors TRI, ANC
Forearm Pronators/Supinators SUP, PT, PQ
Wrist/Finger Flexors FCR, FCU, PL, FDS, FDP, FPL
Wrist/Finger Extensors ECR, ECU, EDC, EDM, EIP, EPL, EPB, APL
Hand Intrinsics OP, RI2, RI3, RI4, RI5, LU_, UI_
Lower Extremity Glutes (Hip Extensors) GMAX, GMED, GMIN, PIR, GEM, OINT, QFM, glmax, glmed, glmin, piri
Quadriceps (Knee Extensors) VAS, RF, VI, VL, VM, vasint, vaslat, vasmed, recfem
Hip Flexors/Adductors ILIACUS, PSOAS, ADD, PECT, GRAC, TFL, iliacus, psoas, add, grac, tfl, sart
Hamstrings (Knee Flexors) BFLH, BFSH, SEMIM, SEMIT, bflh, bfsh, semiten, semimem
Calves (Plantar Flexors) GAS, SOL, soleus, gaslat, gasmed
Other Leg/Foot TA, EDL, EHL, FDL, FHL, TP, PER, edl, ehl, fdl, fhl, perbrev, tibant, tibpost, perlong

A-J Detailed Descriptions of Residual RL Policy

The residual study adapts an imitation prior through a combination of bounded corrections, task objectives and reference timing. Increasing stair-tracking error motivates the intervention but does not identify which component causes failure or recovery.

A-J 1 Action Space and Residual Scaling

Clip the residual output, scale it, add it to the frozen base action and clip the sum to valid action bounds. Generic action-rate and activation-cost penalties are disabled in the shared adaptation setup; a small residual L2 cost with coefficient −0.02 remains. Walking separately reinstates task-specific activation regularization below.

Equation 38
at\displaystyle a_{t} =clip​(abase+α⋅clip​(ares,−1.0,1.0),−1.0,1.0)\displaystyle=\text{clip}\left(a_{\text{base}}+\alpha\cdot\text{clip}(a_{\text{res}},-1.0,1.0),-1.0,1.0\right) (38)

A-J 2 Adaptations for Stair Navigation

Stair tracking uses planar root error, omitting vertical mismatch so height changes do not automatically destroy imitation reward. Reference time advances normally below 0.4 m planar error and at half speed otherwise. A progress reward combines velocity toward a reference pelvis one second ahead with upward velocity. These objective and pacing changes accompany residual learning.

Equations 39–40
droot\displaystyle d_{\text{root}} =‖𝐩x​y−𝐩^x​y‖22\displaystyle=\|\mathbf{p}_{xy}-\mathbf{\hat{p}}_{xy}\|_{2}^{2} (39)
rprog\displaystyle r_{\text{prog}} =wprog​(clip​(vproj,0,2.0)\displaystyle=w_{\text{prog}}\Big(\text{clip}(v_{\text{proj}},0,2.0) (40)
+0.5⋅clip(vz,0,1.5))\displaystyle+0.5\cdot\text{clip}(v_{z},0,1.5)\Big)

A-J 3 Walking Adaptation

Walking uses residual authority 0.2 in stance and 0.6 in swing, gated by foot contact. An activation penalty sums squared activation magnitude and temporal changes with weight 0.05 and smoothing coefficient 0.1 over 416 muscles. This encourages quieter commands without turning activation cost into a direct metabolic measure.

Equations 41–42
αi​(t)\displaystyle\alpha_{i}(t) =αstance+(αswing−αstance)⋅ℐswing​(i,t)\displaystyle=\alpha_{\text{stance}}+(\alpha_{\text{swing}}-\alpha_{\text{stance}})\cdot\mathcal{I}_{\text{swing}}(i,t) (41)
pmet\displaystyle p_{\text{met}} =−wmet∑j=1N(aj,t2+λ(aj,t−aj,t−1)2)\displaystyle=-w_{\text{met}}\sum_{j=1}^{N}\left(a_{j,t}^{2}+\lambda(a_{j,t}-a_{j,t-1})^{2}\right) (42)

A-J 4 Adaptations for High-Impact Locomotion (Run Task)

During running flight, tracking penalties are halved and a forward propulsive reward is enabled. A linear soft penalty applies to vertical ground force beyond 2.5 body weights. This threshold is a training cost, not a hard impact limit: reported peak force after adaptation remains 6.07 body weights.

Equations 43–44
rpropel\displaystyle r_{\text{propel}} =wvel⋅max⁡(0,vx−v^x)⋅ℐair\displaystyle=w_{\text{vel}}\cdot\max(0,v_{x}-\hat{v}_{x})\cdot\mathcal{I}_{\text{air}} (43)
pimpact\displaystyle p_{\text{impact}} =−wgrf⋅max(0,Fz−Flimit)\displaystyle=-w_{\text{grf}}\cdot\max(0,F_{z}-F_{\text{limit}}) (44)

A-K Extended Discussion on Limitations and Future Directions

The extended limitations separate missing task coverage, missing physiological evidence and untested transfer to physical systems.

A-K 1 Absence of Dexterous Manipulation

The tasks emphasize gross movement and basic interactions. Fine grasping, tool use and in-hand manipulation require more detailed hand/tendon modeling and coordinated arm–finger control that this benchmark does not yet evaluate.

A-K 2 Transitioning to Compositional and Long-Horizon Tasks

Most evaluations isolate atomic objectives. Longer sequences require reliable skill transitions, balance management and potentially fatigue modeling. The present results do not establish those compositional capabilities.

A-K 3 Constraints in Comprehensive Physiological Validation

Matched human recordings cover only selected muscles in walking, running and stairs. Normalized phase-optimized correlations cannot validate comprehensive biological similarity. Broader synchronized motion/EMG datasets would be needed for dynamic or unconventional tasks.

A-K 4 Sim-to-Real Gap and Hardware Validation

All results are simulated. Physical tendon-driven bodies introduce uncertain actuation, sensing and communication effects, and transfer to electric humanoids would require another mapping from muscle activity to motor control. Neither deployment route is validated here.

Concept demonstration

Open image ↗