{"id":"6480d07b-fa3e-4dcf-ab8d-3d16225b6c9b","arxiv_id":"2412.08270","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":10,"one_line_summary":"A musculoskeletal humanoid learns a direct neural model from ankle-pedal control input to car velocity and uses gradient-based optimization to track target speeds faster than PID controllers.","lead":"Researchers trained a neural network to control a musculoskeletal humanoid's ankle so it presses a car accelerator pedal to reach a target speed. The robot learned from 60 seconds of random pedal movements and then matched target speeds faster than two PID controllers in single-run experiments.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central generalization claim is untested: DDC-Net is trained on 300 samples from a single random episode, yet no held-out prediction error or repeated closed-loop trials are reported, so the reported convergence times could reflect overfitting or a lucky run.","rationale":"The reader's conditional verdict is appropriate. The paper's method is sensible and the hardware demonstration is interesting, but the central quantitative claim rests on a learned model whose predictive accuracy on unseen states/inputs is never evaluated. My concern is not that the approach is wrong in principle; it is that the reported superiority over PID could be an artifact of a lucky single run or of exploiting DDC-Net's overfitting. Adding a held-out prediction check and repeated closed-loop trials would directly address this. This does not change the reader's conditional verdict.","tokens_in":9222,"tokens_out":6801,"duration_ms":89971,"concrete_test":"Retain a held-out 15 s of the random-data episode and report DDC-Net test MSE (for each horizon 1..30) against training MSE; then run the Proposed controller for 10 independent trials at each of 5 and 10 km/h starting from rest, recording T_conv for each trial and reporting mean±std. If the held-out MSE is high or the trial-to-trial spread is comparable to the reported convergence times, the real-time generalization claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing condition is that the neural network f in Eq. (1), trained on a single 60-second random-control episode (300 samples at 5 Hz), is an accurate enough predictive model of the musculoskeletal-vehicle dynamics for gradient-based optimization to produce physically realizable pedal commands. This condition is untested in the paper: no train/test prediction error for DDC-Net is reported, and no repeated closed-loop trials are shown. With N=30, the network has an input dimension of 34 and a 30-dimensional output but is trained on at most ~270 heavily overlapping windows, so it can overfit the single trajectory. During control, the optimizer may drive u_seq to regions not visited by the random controller (whose increments are only -1 to +2 deg), and any extrapolation error of f directly corrupts the optimized command. The claimed convergence times (0.9 s and 5.3 s) are single observations with no error bars, so they cannot establish that the method reliably realizes the intended task state.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DDC-Net, a fully-connected neural network f that maps the current task state i_t and a time-series of N future control inputs u_{t+1},...,u_{t+N} to a time-series of predicted task states s_{t+1},...,s_{t+N}. The network is trained offline on 300 samples (60 s at 5 Hz) of random ankle-pitch pedal control data collected on the musculoskeletal humanoid Musashi driving a COMS electric vehicle on rollers. In the control phase, the method backpropagates a loss between the predicted and target car-velocity sequences through f to iteratively optimize the control input sequence, then executes the first element at 5 Hz. The paper reports convergence times of 0.9 s (within 20% of 5 km/h) and 5.3 s (within 10% of 10 km/h), claiming faster convergence than two PID-based baselines. The contribution is framed as task-specific self-body controller acquisition that bypasses explicit modeling of the musculoskeletal vehicle-driver dynamics.","tokens_in":9461,"tokens_out":3399,"duration_ms":37990,"significance":"If the central claim were substantiated, the method would be a practically valuable way to acquire controllers for musculoskeletal humanoids in tasks where analytic modeling is intractable: it learns a direct input-to-task-state map from a single short random episode and requires no manual PID tuning. The paper also contains a useful conceptual discussion of control-input/state spaces for musculoskeletal versus axis-driven humanoids. However, the current evidence is limited to single trajectories without repeated trials, and no held-out prediction accuracy of the learned dynamics model is reported. The strength of the claimed improvement over the PID baselines therefore rests on unquantified generalization and run-to-run variability, which the paper does not address.","major_comments":[{"comment":"The training set is a single 60-second Random-control episode (300 samples), and the only validation mentioned is 'we use the 1/5 of the data as test data' with no numerical result reported. No train/test prediction error for DDC-Net is given anywhere, so there is no evidence that f generalizes beyond the single training trajectory. Since N=30 gives an input dimension of 34 and output dimension of 30, and the roughly 270 training windows are heavily overlapping, a 5-layer network could easily overfit this data. Please report the held-out prediction error of DDC-Net and, ideally, evaluate the predicted state trajectories on a separately collected random-control episode or on a portion of the control range not used in training.","section":"III-B and IV-C"},{"comment":"The headline convergence times (0.9 s for 5 km/h within 20%, 5.3 s for 10 km/h within 10%) are single observations from what appears to be a single experimental run for each target. The PID1 and PID2 results in Figs. 7 and 8 are also single trajectories. Without repeated trials, error bars, or another measure of variability, the claimed improvement over PID1/PID2 cannot be distinguished from run-to-run variation. At minimum, please run several repetitions of each condition (Proposed, PID1, PID2) and report the mean and spread of T_conv for each criterion.","section":"IV-C and Fig. 10"},{"comment":"The convergence criterion is asymmetric and chosen post hoc: A=20% for v_target=5 km/h but A=10% for v_target=10 km/h, with the stated justification that tracking 5 km/h is difficult. This makes the comparison between Proposed and PID1/PID2 not apples-to-apples: PID2's 10.7 s to the 20% threshold at 5 km/h is compared with Proposed's 0.9 s to the same threshold, while at 10 km/h the threshold changes to 10%. Please use a common threshold (or report both thresholds for every method) and fix the evaluation criterion before comparing methods.","section":"IV-B"},{"comment":"The optimization procedure can produce control inputs far outside the distribution of the random-control training data. The random controller perturbs u by only -1 to +2 degrees per step (Section IV-C), while the first-stage optimization uses beta1=3.0 degrees for N1=10 iterations with a batch of noisy initializations, and the only constraint on u is clamping to [0,50] degrees. Since the accuracy of f is never evaluated outside the observed input range, the optimized u_seq may exploit extrapolation error rather than true dynamics. Please provide an analysis or experiment showing that the optimized control inputs remain within (or are robust beyond) the range spanned by the training data, for example by plotting the distribution of optimized u_t during the control phase against the training distribution.","section":"III-C"}],"minor_comments":[{"comment":"The adjacent-error term in Eq. (2) is written with u_init^seq, although the quantity being optimized is the target control input sequence; using the same symbol for the initial guess and the optimized variable is confusing.","section":"III-C, Eq. (2)"},{"comment":"The caption of Fig. 9 appears truncated ('Target joint angle of ankle-p'), and the target velocity used in the Random controller of Eq. (11)-(12) is not specified; please state whether the data-collection episode targeted 5 km/h, 10 km/h, or a time-varying target.","section":"Fig. 9"},{"comment":"The random controller equations are referenced in the text as Eq. (11) and Eq. (12) but are not numbered in the manuscript; please number them.","section":"IV-C"},{"comment":"The statement 'we only had to tune alpha after setting all parameters' is inconsistent with the many hand-chosen hyperparameters (N, hidden-layer sizes, beta1, beta2, Nbatch, N1, N2, delta_ubatch). A sensitivity study of at least alpha and beta would support the claim that the method is easy to deploy.","section":"V"}],"recommendation":"major_revision","confidential_remarks":"The paper reports an interesting hardware demonstration, but the central claim of fast, reliable task-state realization is under-evidenced because of single-run comparisons and the absence of any held-out evaluation of the learned dynamics model. These issues are fixable within the manuscript's scope by adding repeated trials and prediction-error results. I would also gently push the authors to address the asymmetric convergence thresholds, which currently make the comparison with PID methods look potentially unfair even if unintentionally so."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a well-scoped idea. The authors train a neural network (DDC-Net) on a single 60-second random-control episode to map current state plus a time-series of ankle-pitch commands to a time-series of car velocities, then run gradient-based MPC through that network at 5 Hz. For a musculoskeletal humanoid, where the flexible body is hard to model, learning a direct task-level model is a sensible move, and the experimental setup (COMS car, Musashi) is appropriate. The architecture also fixes a real gap in the authors' earlier torque-sequence work [15] by outputting time-series task states and regularizing the input sequence.\n\nThe paper is clearly written and honest: the discussion acknowledges that PID gains could be better tuned, that high gains cause vibration, and that scaling to multi-DOF input is open. That counts.\n\nWhere I land on the weaknesses: the reader's skeptics are basically right, though I'd soften the stress-test conclusion from \"overfitting or a lucky run\" to \"we cannot yet distinguish.\" Yes, they use 300 samples from one random episode, and the network has 34 inputs and 30 outputs, so overfitting is a real risk. But they do hold out 1/5 of the data for model selection, and the closed-loop behavior on untrained target velocities is at least an out-of-sample test. The bigger problem is that each condition is a single run with no error bars, and the convergence thresholds differ between conditions (20% for 5 km/h, 10% for 10 km/h) with the asymmetry justified after seeing results. That makes the headline numbers (0.9 s vs 22.9 s, 5.3 s vs 27.5 s) much less compelling than they look. We simply cannot tell whether DDC-Net reliably generalizes or whether these are lucky draws.\n\nAlso, the input sequence during random collection uses increments between -1 and +2 deg, but the MPC optimizer can propose much larger changes; if DDC-Net extrapolates badly there, the optimized command is meaningless. The paper doesn't report how often the optimizer respects the training distribution. That's a testable, fixable omission.\n\nIn proportion: the central idea is sound, the demonstration is plausible, and the flaws are evidence-quality, not conceptual. This is a poster-to-short-paper-level contribution, not a slam dunk, but it deserves a serious referee rather than a desk reject. I'd ask for repeated runs, consistent threshold definitions, a held-out prediction loss number, and ideally code/data. If those come back clean, this becomes a solid reference for task-level learning on high-DOF soft robots.\n\nFor our group: worth a read for anyone doing learned dynamics MPC, but not urgent. Recommendation: send to peer review with the expectation of heavy revision around experimental rigor.","headline":"A credible but under-evidenced application of learned-dynamics MPC to musculoskeletal pedal control; the central generalization claim is plausible but not yet demonstrated.","tokens_in":10034,"tokens_out":3122,"would_cite":false,"duration_ms":34895,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims a musculoskeletal humanoid can learn a pedal controller directly from 60 seconds of random driving data, reaching target speeds faster than PID.","keywords":["musculoskeletal humanoid","DDC-Net","task-specific control","neural network dynamics model","autonomous driving","pedal control","real-time control","backpropagation to input"],"falsifier":"Run the learned DDC-Net controller on ten fresh trials from different initial velocities (for example, starting at 2 km/h and 7 km/h instead of 0) and compare the predicted velocity trajectory against the measured one over the six-second horizon; if predicted and actual velocities diverge beyond the reported 10-20% bands, or if the commanded pedal sequence fails to converge faster than PID2 on those starts, the claim that 60 seconds of random data suffice for real-time control is refuted.","tokens_in":9037,"feed_emoji":"🚗","tokens_out":5476,"duration_ms":54584,"temperature":0.7,"pith_summary":"This paper tries to show that a musculoskeletal humanoid can acquire a task-specific controller without building an accurate model of its own flexible body. The proposed method, DDC-Net, learns a direct time-series map from ankle-pitch pedal commands to future car velocity from 300 random driving samples collected over 60 seconds, then inverts that map in real time by backpropagating the target-velocity error into the command sequence. In accelerator-pedal experiments the learned controller reaches 5 km/h within 20% in 0.9 seconds and 10 km/h within 10% in 5.3 seconds, beating two PID baselines at those thresholds. The point is that the robot does not need an accurate joint-muscle model; it needs only a task-level input-output model and a way to optimize through it.","feed_headline":"A robot learns pedal control from 60 seconds of random data","feed_subtitle":"A musculoskeletal humanoid's learned network reaches target speeds faster than tuned PID controllers.","key_machinery":"The central object is DDC-Net, a five-layer fully connected neural network implementing Eq. (1): $f$ maps the current state $i_t$ together with $N$ future control inputs $u_{t+1},\\dots,u_{t+N}$ to $N$ predicted future task states $s_{t+1},\\dots,s_{t+N}$. Each hidden layer is batch-normalized with sigmoid activations, with hidden units $\\{80,50,20\\}$, input dimension $N_i + N \\times N_u = 4 + 30$, and output dimension $N \\times N_s = 30$. It carries the argument because it replaces the unmodelable chain of muscle wires, soft tissue, and vehicle dynamics with a differentiable surrogate that can be trained from random motion data and then inverted by gradient descent on the control input. The control loop's loss $L = \\mathrm{MSE}(s_{\\text{pred}}, s_{\\text{target}}) + \\alpha \\,\\mathrm{AdjacentError}(u_{\\text{init}})$ makes the inversion produce smooth command sequences, and the two-stage batched initialization ($N_{\\text{batch}}=10$ noisy copies, $N_1=10$ then $N_2=20$ gradient steps) keeps the real-time optimization stable at 5 Hz.","core_discovery":"On the paper's own terms, the central discovery is that for under-modeled musculoskeletal robots the right object to learn is not the joint-muscle mapping but the dynamic relationship between the control input sequence and the task state sequence. DDC-Net is defined by Eq. (1): a function $f$ that takes the current state $i_t$ (car velocity, car acceleration, ankle-pitch angle, ankle-pitch velocity) and a length-$N=30$ sequence of future ankle-pitch angles $u_{t+1},\\dots,u_{t+N}$ and outputs the predicted velocities $s_{t+1},\\dots,s_{t+N}$. After training on random pedal inputs at 5 Hz for 60 seconds, the network is used as a differentiable forward model: the control phase initializes a command sequence from the previous solution, adds noise to form a batch, minimizes the MSE between predicted and target velocity plus a smoothness penalty on adjacent commands, and backpropagates the gradient into the input sequence. The first optimized command is sent to the robot; repeated at 5 Hz, this realizes target speeds in roughly one to five seconds. The paper reports that this converges faster than PID controls and, unlike them, requires no gain tuning for the pedal task.","pith_inferences":["The algorithm is, in effect, a learned model-predictive controller with a six-second horizon and a differentiable forward model; any plant whose dynamics are hard to model analytically but easy to excite with random inputs could use the same recipe, not just musculoskeletal humanoids.","Because training data come from one random controller run near the target region, the learned model is only guaranteed where that run explored; an active data-collection scheme that re-runs random inputs after seeing large prediction errors could widen the reliable operating envelope.","The batched noisy initialization and normalized gradient descent resemble a tiny population-based optimizer; treating it that way suggests testable variations, such as covariance-matrix-style updates or longer horizons, that might improve convergence further.","A direct comparison with a learned dynamics model plus standard model-predictive control would isolate whether the contribution is the network structure or the backprop-through-input optimizer."],"forward_implications":["Applied to the same robot, the method should control the brake pedal, steering, and body-tilt compensation without retuning the network, using the same DDC-Net structure.","The control input need not be joint angle; the same loss-and-backprop scheme should work with joint velocity, joint torque, or muscle tension commands.","Because the target state sequence enters the loss, the method should track time-varying velocity profiles, such as smooth stopping trajectories chosen by jerk minimization.","The method removes manual PID gain tuning for pedal control; the main remaining tuning parameter is $\\alpha$, the smoothness weight.","Scaling to multi-DOF tasks requires reducing the control input space, for example by muscle synergy or autoencoder compression, because the input space grows exponentially with the number of control dimensions."],"supporting_citations":[{"why":"Describes the Musashi musculoskeletal humanoid platform whose ankle joint is used in the experiments.","marker":"[9]"},{"why":"Prior online joint-muscle mapping method; the paper argues it cannot realize task-space goals directly.","marker":"[7]"},{"why":"Prior online self-body image acquisition; also used here to convert joint-angle commands into muscle lengths.","marker":"[8]"},{"why":"Offline muscle-joint state mapping that handles static states only; motivates the time-series task-state approach.","marker":"[12]"},{"why":"Prior dynamic manipulation with a deep network that lacks time-series task-state output and control-input regularization; the proposed loss addresses both.","marker":"[15]"},{"why":"Reports human pedal feedback bandwidth around 3.5 Hz, used to justify the 5 Hz control rate.","marker":"[17]"},{"why":"Batch normalization used in the hidden layers of DDC-Net.","marker":"[18]"},{"why":"Adam optimizer used to train DDC-Net.","marker":"[19]"}],"fun_headline_variants":["From random data to precise pedal control in under 5 seconds","Direct dynamics learning beats tuned PID for musculoskeletal pedal control","Musculoskeletal humanoid learns pedal control without joint-muscle model","60 seconds of random data teaches robot to hit target speeds","Learn the input-output dynamics, not the body model: robot pedal control"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that 300 random pedal samples taken over one 60-second run capture the robot-vehicle dynamics accurately enough that gradient descent through the learned network produces correct commands on later runs, including speeds and initial conditions not explicitly in the training set; the paper's reported experiments are single runs with no repetition or off-distribution test.","fun_headline_variants_meta":{"raw":{"variants":["From random data to precise pedal control in under 5 seconds","Direct dynamics learning beats tuned PID for musculoskeletal pedal control","Musculoskeletal humanoid learns pedal control without joint-muscle model","60 seconds of random data teaches robot to hit target speeds","Learn the input-output dynamics, not the body model: robot pedal control"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000463,"raw_usage":{"total_tokens":2295,"prompt_tokens":905,"completion_tokens":1390,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":521,"completion_tokens_details":{"reasoning_tokens":1304}},"tokens_in":521,"tokens_out":1390,"duration_ms":11003,"temperature":1.0,"reasoning_tokens":1304,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T18:00:23.846293+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the learned DDC-Net controller on ten fresh trials from different initial velocities (for example, starting at 2 km/h and 7 km/h instead of 0) and compare the predicted velocity trajectory against the measured one over the six-second horizon; if predicted and actual velocities diverge beyond the reported 10-20% bands, or if the commanded pedal sequence fails to converge faster than PID2 on those starts, the claim that 60 seconds of random data suffice for real-time control is refuted.","supporting_citations":[{"cited_title":"Online Self-body Image Acquisition Con- sidering Changes in Muscle Routes Caused by Softness of Body Tissue for Tendon-driven Musculoskeletal Humanoids,","cited_arxiv_id":null,"evidence_quote":"Prior online self-body image acquisition; also used here to convert joint-angle commands into muscle lengths."},{"cited_title":"Learning Nonlinear Muscle-Joint State Mapping Toward Geometric Model-Free Tendon Driven Musculoskeletal Robots,","cited_arxiv_id":null,"evidence_quote":"Offline muscle-joint state mapping that handles static states only; motivates the time-series task-state approach."},{"cited_title":"Dynamic Manipulation of Flexible Objects with Torque Sequence Using a Deep Neural Network,","cited_arxiv_id":null,"evidence_quote":"Prior dynamic manipulation with a deep network that lacks time-series task-state output and control-input regularization; the proposed loss addresses both."},{"cited_title":"Steer-by-wire: a study into the bandwidth and force requirements,","cited_arxiv_id":null,"evidence_quote":"Reports human pedal feedback bandwidth around 3.5 Hz, used to justify the 5 Hz control rate."},{"cited_title":"Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift,","cited_arxiv_id":null,"evidence_quote":"Batch normalization used in the hidden layers of DDC-Net."}],"review_version":1}