{"id":"ac3122b0-cfaa-4489-8eee-9b1a98803e9f","arxiv_id":"2505.08238","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"MPC^2 controls a 700-muscle simulated human body without training by planning target postures with sampling-based MPC and coordinating muscles with a morphology-aware proportional controller.","lead":"This paper introduces a control algorithm that makes a simulated full-body human model with 700 muscles stand, walk, climb stairs, and kick a ball without any training, using model-based planning instead of reinforcement learning. The method works in near real time and could speed up biomechanics research and robot control design.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (10) is the sole low-level inversion from planned posture to muscle control, yet it is un-derived and inverts the simplified model in Eq. (2) rather than the MuJoCo dynamics in Appendix A; one-step tracking is untested.","rationale":"The reader identifies the right structural location: Eq. (10) is the only mechanism that converts planned postures into muscle activations, and it is asserted without derivation. However, I do not fully endorse the dimensional objection as stated, because the plain-text rendering is ambiguous; if the intended formula is u^* = a + \\tau_2(a^*-a)/(\\Delta t - \\tau_1(a^*-a)), it is dimensionally consistent and follows from a one-step solve of Eq. (2). The stronger and less parsing-dependent problem is that Eq. (2) is not the activation dynamics actually used in MuJoCo: Appendix A specifies sign-dependent \\tau(u,a) with \\tau_act and \\tau_deact, not the constant-coefficient form (u-a)\\tau_1+\\tau_2. Unless \\tau_1,\\tau_2 and the integration scheme are chosen to make Eq. (10) an exact inverse of the Appendix A model, the low-level tracking assumption is unverified. This is load-bearing because the empirical videos and Table 1 demonstrate that some implementation works, but the published equations alone do not let a reader reproduce or verify the crucial low-level law. I would not reject: the qualitative evidence is consistent and the closed-loop nature of the controller could compensate for an approximate inverse. But the missing derivation, missing parameter values, and absence of tracking-error data justify keeping the reader's conditional verdict. A one-step activation-tracking test against the actual MuJoCo muscle model would settle whether this concern lands.","tokens_in":13204,"tokens_out":14004,"duration_ms":149751,"concrete_test":"Standalone one-step check: in the actual MuJoCo MS-Human-700 muscle model (Appendix A), for a grid of (a, a^*, \\Delta t) pairs spanning the activation range [0,1], compute u^* from Eq. (10) using the authors' \\tau_1,\\tau_2 (or default variants if unreported), clip u^* to [0,1], simulate one step, and measure |a_next - a^*|. If the median absolute tracking error is not small (say >0.1), Eq. (10) is not the reliable one-step inverse the paper assumes, and the hierarchical controller would need explicit closed-loop compensation evidence.","verdict_should_be":"UNCHANGED","load_bearing_attack":"MPC^2's low level reduces to Eq. (10), which converts a desired actuator force into a control u^* in one step; if this inversion is wrong or not matched to the simulator's activation dynamics, the high-level posture plan has no other mechanism to reach the planned posture. The paper says 'we are able to derive' Eq. (10) but gives no derivation. Taking Eq. (2) as \\dot a = (u-a)/((u-a)\\tau_1+\\tau_2), a one-step solve gives u^* = a + \\tau_2(a^*-a)/(\\Delta t - \\tau_1(a^*-a)); the printed rendering 'u^* = a + \\tau_2(a^*-a)/\\Delta t - \\tau_1(a^*-a)' is dimensionally inconsistent unless the entire \\Delta t-\\tau_1(a^*-a) is the denominator. More importantly, the actual MuJoCo muscle dynamics in Appendix A are \\dot a = (u-a)/\\tau(u,a) with sign-dependent \\tau(u,a) = \\tau_act(0.5+1.5a) for u>a and \\tau_deact/(0.5+1.5a) for u\\le a, which is not the simplified model in Eq. (2). Thus Eq. (10) is not an exact inverse of the simulator's activation dynamics. The paper does not report \\tau_1,\\tau_2, does not state how u^* is clipped to [0,1], and gives no activation-tracking error data. Because this inversion is the only bridge from the 37-dimensional posture plan to the 700 muscle controls, the central zero-shot claim depends on an untested and under-specified low-level law.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MPC^2, a hierarchical model-based controller for high-dimensional musculoskeletal systems. A high-level sampling-based MPC (MPPI with 'instant rollouts') plans a sequence of target postures in a low-dimensional joint space (37 dimensions for the 700-muscle human model), while a morphology-aware proportional controller maps the target posture to per-muscle activation commands using Jacobian-scaled gains. The authors report zero-shot standing, walking over flat/rough/sloped/stair terrains, sports imitation, and control of an ostrich and an arm model, all without training, plus automated cost-function tuning via Bayesian optimization. Ablations support the instant-rollout and morphology-aware-gain components, and comparisons are made against DRL (DynSyn) and various MPC baselines.","tokens_in":13636,"tokens_out":7972,"duration_ms":79367,"significance":"If the empirical claims hold, this is a substantial advance: a training-free, model-based controller matching or exceeding the task coverage of DRL methods that require days of training on a 700-actuator musculoskeletal model. The hierarchical decomposition is clean and the ablation study gives evidence for both main components. The black-box cost-function optimization is a practical and timely contribution. However, the paper's strongest claims rest on the low-level actuator inversion in Eq. (10), which is under-derived and untested against the actual MuJoCo muscle dynamics, and the flagship empirical results are not backed by quantitative metrics for most tasks. With the requested clarifications and added evidence, the contribution would merit publication.","major_comments":[{"comment":"Equation (10) is the only mechanism that converts the planned high-level posture into muscle activations, yet it is introduced with 'we are able to derive' and no derivation is given. As printed, the expression is ambiguous: a literal reading of 'u* = a + τ2(a*−a) Δt − τ1(a*−a)' is dimensionally inconsistent (the first fractional term is dimensionless while τ1(a*−a) has units of time). If the intended formula is u* = a + τ2(a*−a)/(Δt − τ1(a*−a)), please derive it from Eq. (2) and state the assumptions. More substantively, the actual MuJoCo activation dynamics in Appendix A use τ(u,a) = τ_act(0.5+1.5a) for u>a and τ_deact/(0.5+1.5a) for u≤a, which is not the simplified form in Eq. (2); the paper does not report the values of τ1 and τ2, does not specify how u* is clipped to [0,1], and provides no activation-tracking error data. Because this inverse law is the sole bridge between the 37-dimensional posture plan and the 700 muscle controls, the central zero-shot claim depends on an untested and underspecified low-level law. Please add the derivation, the exact parameter values and clipping, and a one-step tracking experiment, or show that residual tracking errors are absorbed by the high-level MPC loop.","section":"§4.2, Eq. (10)"},{"comment":"The definition of the gain vector K is ambiguous. In the text K is described as a per-actuator vector (k1,...,k_du), but the right-hand side of Eq. (11) as written appears to be a scalar (a sum over joints of a product involving the Jacobian column). If col_i(J_m) is a vector over actuators, then the notation 'col_i(J_m) · [z*_i − M_pos(s_t)_i]' mixes a vector with a scalar; please make explicit how each per-actuator gain k_j is extracted from the Jacobian. In addition, the scalar k̄ is the only free low-level controller parameter, but its value is never reported. The paper should state the chosen k̄ and provide a sensitivity analysis to show that the results do not hinge on a narrow range of this parameter.","section":"§4.2, Eq. (11)"},{"comment":"The central empirical claim is stable zero-shot control across Stand, Walk, Rough, Slope, and Stair, but only qualitative statements and videos are provided for most tasks. No numerical success criteria, failure counts, or distance/speed metrics are reported for the terrain tasks, and the performance metric plotted in Figure 5 is not defined in the text. Table 1 reports walking distance for MPC^2 and six MPC baselines, but this table is not analyzed in the main text. Please define the evaluation metrics, report means and standard errors over multiple seeds for every task, and state whether 'stable control' is a thresholded criterion. Without this, the flagship claim cannot be verified quantitatively.","section":"§5.1 and Table 1"},{"comment":"The abstract and contribution list describe the method as achieving 'near-real-time' control, but the reported simulation conditions are 20% of real-time speed in the main thread (10% for Stair and 1% for soccer), meaning the method is 5–100 times slower than real time depending on the task. The claim that 'control sequences to complete the task can be generated within 2 minutes' is not equivalent to near-real-time if the simulated task itself takes 50 seconds or more at 20% speed. Please either revise the wording to reflect the actual wall-clock latency per control step or report the per-step planning latency separately from the simulation speed.","section":"§5.1 and Abstract"}],"minor_comments":[{"comment":"The statement that the target actuator length l* 'can be computed with model forward dynamics' is imprecise; actuator length is a kinematic function of joint positions, not of forward dynamics. Please use the term 'forward kinematics' or 'muscle path geometry'.","section":"§4.2"},{"comment":"In the Walk task definition, the cost is labeled 'Cstand' rather than 'Cwalk', and the expression 'Cstand = 100(Cheight + Cupright + Cbalance + 10Cvf + 10Cvdir + 100Cbf + 5Cjp + 50Cfc' is missing a closing parenthesis; the same missing parenthesis appears in several other task definitions.","section":"Appendix B"},{"comment":"The normalization 'a = 1/(1 + e^{−5(a−0.5)})' uses the symbol a on both sides; presumably the left-hand side is the normalized control input and the right-hand side uses the raw policy output. Please clarify the notation, since as written it is self-referential.","section":"Appendix A"},{"comment":"The notation 'R_MP(z, H)' in Algorithm 1 and Eq. (8) is not explicitly defined; please state that it denotes the H-step cumulative cost under the low-level policy π_MP.","section":"Algorithm 1"},{"comment":"There is a typo in the opening sentence: 'morphology-aware proposition controller' should be 'morphology-aware proportional controller'.","section":"§5.4"},{"comment":"The axes of Figure 5(a) and 5(b) are not labeled in the text or caption; please specify what quantity is plotted against clock time (e.g., forward displacement, balance cost, or walking speed).","section":"Figure 5"},{"comment":"The phrase 'Following the implementation in sh MPC' appears to reference a baseline or prior work without a citation; please provide the reference or remove the incomplete name.","section":"Appendix C.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is already published at ICLR 2025; if this is a journal version, the editor may wish to consider whether the added material goes beyond the conference version. The main technical risk is the low-level inverse actuator law, which as written is not an exact inverse of the MuJoCo activation dynamics and lacks any tracking-error validation; I would not recommend acceptance until that is addressed. The absence of code or a documented reproducibility package is also a concern for a systems paper of this type, though the project page videos help."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: MPC^2 is a genuine empirical advance — training-free MPC controlling a 700-muscle human model for standing, walking, slopes, stairs, and sports imitation — but the paper's low-level actuator inversion (Eq. 10) is un-derived and poorly presented, and the speed claims are overstated. I'd take the zero-shot results seriously, but I'd want the low-level law clarified and validated before citing the mechanism as fact.\n\nWhat's actually new: the hierarchical recipe — sampling-based MPC over 37 posture dimensions plus a Jacobian-scaled proportional controller that maps posture error to 700 muscle commands — looks sensible, and the ablations earn their keep. Instant rollouts help in the lean-back recovery test, and morphology-aware gains beat constant-gain and PD variants. The terrain, perturbation, muscle-failure, ostrich, and arm results are a broad empirical package, and the 2-minute versus day-long comparison with DynSyn is a real advantage for reward-function iteration.\n\nThe soft spots are concentrated around Eq. (10). As printed, it is dimensionally ambiguous; the intended expression is almost certainly u* = a + τ2(a*-a)/(Δt − τ1(a*-a)), and the paper gives no derivation. More substantively, the derivation uses the simplified activation dynamics of Eq. (2), not the MuJoCo muscle model in Appendix A, which has a sign-dependent time constant. The paper never reports activation-tracking error, so the one bridge from planned posture to 700 muscle activations is unverified. That is a genuine gap, not a nitpick. The near-real-time language is also optimistic given the main thread runs at 20% real-time speed (1% for soccer), and the 'no previous methods' claim would need a more systematic comparison to back. No code or data accompanies the paper.\n\nThe central empirical claim — that this controller stabilizes a high-dimensional musculoskeletal model zero-shot — is plausible and consistent with the videos and quantitative walk distances. The missing derivation and tracking validation are addressable. This paper deserves serious peer review; I would send it out, and ask the authors to derive Eq. (10) properly, show activation tracking, and recalibrate the speed claims.\n\nIf you're deciding whether to engage: if you work in musculoskeletal control or model-based RL, yes, read it and probably cite the empirical result. Just don't rely on the low-level inverse law as stated.","headline":"A real zero-shot MPC result for 700-muscle control, but the low-level inverse law is under-derived and the speed claim overstates.","tokens_in":14100,"tokens_out":3931,"would_cite":true,"duration_ms":36349,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Zero-shot MPC controls a 700-muscle whole-body human model without any training.","keywords":["model predictive control","musculoskeletal systems","zero-shot control","hierarchical control","sampling-based MPC","whole-body humanoid","cost function optimization","overactuated systems"],"falsifier":"Replace the one-step inverse law (Eq. (10)) with a numerically stable inversion of the first-order activation ODE (for example an implicit Euler solve), keeping every other component fixed. If walking distance collapses toward the MPPI baseline or stand time drops sharply, then the single-step inversion is load-bearing. A direct check is to log per-muscle activation error over a ten-second stand: if most actuators miss their target activation by more than a few percent each step, the one-step inversion is not actually holding.","tokens_in":13021,"feed_emoji":"🦵","tokens_out":4718,"duration_ms":46906,"temperature":0.7,"pith_summary":"This paper claims that high-dimensional musculoskeletal control does not require reinforcement learning: a hierarchical, model-based planner can achieve stable whole-body motion in minutes instead of days. The proposed $MPC^{2}$ splits the problem into high-level posture planning and low-level muscle coordination. A sampling-based model predictive controller plans a 37-dimensional target posture, and a morphology-aware proportional controller translates that posture into 700 individual muscle commands. The authors demonstrate stable zero-shot standing, walking on flat, rough, sloped, and stair terrain, and sports imitation on a whole-body human model, where state-of-the-art DRL baselines need at least a day to train. The fast planning loop also enables automatic cost-function tuning by black-box optimization, reducing human reward engineering.","feed_headline":"Zero-shot MPC gets a 700-muscle human model walking","feed_subtitle":"Two-level planning replaces days of reinforcement learning with minutes of sampling-based control.","key_machinery":"The load-bearing mechanism is the two-level hierarchy: a model predictive position controller that samples and optimizes a 37-dimensional target posture (the major joint coordinates), and a morphology-aware proportional controller that computes target muscle lengths, converts them to actuator forces, and inverts the first-order activation dynamics to produce neural-excitation commands. The proportional gains are not constant; they are scaled by the absolute value of the Jacobian columns relating each actuator to the planned posture change, as in Eq. (11). This Jacobian-based gain schedule lets the low level automatically prioritize the muscles most relevant to the current posture and motion phase, which the paper identifies as the reason the high-level planner can operate in a drastically reduced action space.","core_discovery":"$MPC^{2}$ achieves stable zero-shot control of the 700-actuator MS-Human-700 whole-body musculoskeletal model across standing, terrain walking, and sports imitation, without any training procedure. The method plans only a low-dimensional target posture, then uses a proportional controller whose per-muscle gains are set from the muscle Jacobian, so that 700 actuators are coordinated implicitly rather than individually commanded. The paper further shows that this controller continues to walk when major muscles in one leg are suddenly disabled, tolerates large external perturbations, transfers zero-shot to an ostrich musculoskeletal model with the same cost function, and produces a walking gait with over 75% lower summed muscle activation than a DRL baseline. The central empirical claim is that hierarchical model-based planning can replace day-scale learning with minute-scale, near-real-time control in over-actuated systems.","pith_inferences":["The 37-dimensional posture abstraction suggests a general recipe for other high-dimensional over-actuated systems: find a task-relevant low-dimensional manifold (not necessarily joint positions) and a Jacobian-style map from actuators to that manifold.","If the Jacobian-based gain schedule is the core ingredient, a cheaper approximation using only posture-dependent muscle moment arms might suffice on physical hardware where exact muscle Jacobians are uncertain.","The near-real-time claim depends on fast parallel rollouts in simulation; transferring to real hardware would likely require replacing the exact dynamics with a learned model, and robustness under model error could degrade.","The large reduction in summed muscle activation compared with the DRL baseline hints that learned muscle coordination may be over-activating; this is testable experimentally by comparing EMG-like activation sums in the produced gaits."],"forward_implications":["Control generation for new tasks drops from at least a day of DRL training to about two minutes, so cost functions can be iterated automatically via Bayesian optimization within hours.","The same controller and cost function transfer zero-shot across morphologies, demonstrated on the ostrich model, suggesting that the hierarchy captures something morphology-invariant.","Over-actuation becomes a robustness feature: sudden loss of key muscles is absorbed by the remaining actuators without retraining.","Because the planner is training-free, the method could be deployed in settings where tasks and cost functions change frequently, where DRL's training cost would be prohibitive."],"supporting_citations":[{"why":"Supplies the MS-Human-700 whole-body musculoskeletal model that the central control experiments run on.","marker":"Zuo et al., 2024"},{"why":"Provides the MuJoCo physics engine and the first-order actuator dynamics that the low-level controller inverts.","marker":"Todorov et al., 2012"},{"why":"Offers the MJPC platform used for asynchronous sampling-based model predictive control and the framework for real-time control generation.","marker":"Howell et al., 2022"},{"why":"Defines the MPPI sampling update that the high-level posture planner is built on and also serves as a baseline that fails at walking.","marker":"Williams et al., 2016"},{"why":"Is the DynSyn DRL baseline that requires about a day of training and is the main reference point for the time-efficiency comparison.","marker":"He et al., 2024"},{"why":"Underlies the muscle-tendon force-length and force-velocity dynamics used in the actuator model.","marker":"Millard et al., 2013"}],"fun_headline_variants":["Zero-shot MPC walks a 700-muscle human model","700 muscles, zero training: hierarchical MPC walks","Hierarchical MPC: no learning, 700 muscles, walks zero-shot","Minutes not days: hierarchical MPC walks 700 muscles"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The controller can reliably compute, in a single control step, the neural excitation that makes each muscle produce the commanded force, even though the printed inverse formula in Eq. (10) is dimensionally inconsistent and no derivation is shown.","fun_headline_variants_meta":{"raw":{"variants":["Zero-shot MPC walks a 700-muscle human model","700 muscles, zero training: hierarchical MPC walks","Hierarchical MPC: no learning, 700 muscles, walks zero-shot","Minutes not days: hierarchical MPC walks 700 muscles"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000866,"raw_usage":{"total_tokens":3720,"prompt_tokens":881,"completion_tokens":2839,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":497,"completion_tokens_details":{"reasoning_tokens":2771}},"tokens_in":497,"tokens_out":2839,"duration_ms":21243,"temperature":1.0,"reasoning_tokens":2771,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:59:58.698686+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Replace the one-step inverse law (Eq. (10)) with a numerically stable inversion of the first-order activation ODE (for example an implicit Euler solve), keeping every other component fixed. If walking distance collapses toward the MPPI baseline or stand time drops sharply, then the single-step inversion is load-bearing. A direct check is to log per-muscle activation error over a ten-second stand: if most actuators miss their target activation by more than a few percent each step, the one-step inversion is not actually holding.","supporting_citations":[{"cited_title":"Self model for embodied intelligence: Mod- eling full-body human musculoskeletal system and locomotion control with hierarchical low- dimensional representation","cited_arxiv_id":null,"evidence_quote":"Supplies the MS-Human-700 whole-body musculoskeletal model that the central control experiments run on."},{"cited_title":"Mujoco: A physics engine for model-based control","cited_arxiv_id":null,"evidence_quote":"Provides the MuJoCo physics engine and the first-order actuator dynamics that the low-level controller inverts."}],"review_version":1}