PIMBS: Efficient Body Schema Learning for Musculoskeletal Humanoids with Physics-Informed Neural Networks

Kei Okada; Keita Yoneda; Kento Kawaharazuka; Takahiro Hattori

REVIEW 4 major objections 5 minor 17 references

PIMBS: Efficient Body Schema Learning for Musculoskeletal Humanoids with Physics-Informed Neural Networks

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read PIMBS adds physics constraints to body-schema networks so musculoskeletal robots learn muscle maps from very few samples.

desk verdict Real-world claim rests on test-set alpha selection; simulation is strong and the idea is sensible — worth reviewing with a required validation protocol. read the letter →

arxiv 2506.20343 v2 pith:SLH7WS3J submitted 2025-06-25 cs.RO

Kento Kawaharazuka , Takahiro Hattori , Keita Yoneda , Kei Okada This is my paper

classification cs.RO

keywords physics-informedneuralnetworksbodyschemamusculoskeletalhumanoidmuscleJacobiansmall-datalearninglengthestimationtendon-drivenrobots

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Musculoskeletal robots have muscles whose paths deviate from geometric models, so they must learn the mapping from joint angles (and muscle tension) to muscle length from data collected on the body. Collecting enough data is slow, so this paper asks whether physical knowledge can substitute for data. The proposed method, PIMBS, trains the same kind of network as before but adds two losses: one enforcing the statics relation between joint torque and muscle tension through the network's derivative (the muscle Jacobian), and one pinning muscle length to zero at the neutral pose. In a 2-DOF 4-muscle simulation the combined losses reduce held-out error by roughly 60–70% at 3–5 training points for the angle-length map and 20–35% for the angle-tension-length map, with the gap shrinking as data grows. On a real 5-DOF 10-muscle arm the same method only helps after the physics loss is down-weighted, because friction breaks the assumed torque-tension law; with that tuning it again beats the plain-data baseline at small data sizes.

What carries the argument

The load-bearing identity is the static torque-tension relation $\tau = -G^{\mathsf T} f$, where $G$ is the muscle Jacobian obtained by differentiating the learned network. Because the network's derivative is available by backpropagation, the physics loss $L_{\mathrm{pinn}}$ can push the learned map to be consistent with the robot's gravity-compensation torque without any extra measurements; $L_{\mathrm{const}}$ stabilizes the origin. The paper's experiments compare four loss configurations (Basic, Basic+Const, Basic+PINN, Basic+Const+PINN) across simulated and real musculoskeletal structures.

What would settle it

Run the 2-DOF simulation at $N_{\mathrm{train}} = 5$ with a deliberately incorrect joint axis inserted into the computation of $\tau(\theta)$: if Basic+Const+PINN no longer beats Basic+Const by the reported margin, the gain is tied to the joint model being exact; alternatively, add direction-dependent friction to the simulation and check that high $\alpha$ degrades performance as it does on the real robot.

Watch

Extended reading notes

Core claim

PIMBS claims that a physics-informed loss makes body-schema learning more sample-efficient. The network $h$ maps joint angle $\theta$ (and muscle tension $f$) to muscle length $l$; its derivative $G = \partial h/\partial \theta$ gives the muscle Jacobian. The paper adds $L_{\mathrm{pinn}} = \frac{1}{N_{\mathrm{train}}} \sum (G^{\mathsf T} f + \tau)^2$, which encodes $\tau = -G^{\mathsf T} f$ under the assumption that the joint geometry is correct, and $L_{\mathrm{const}} = h(0,0)^2$, which anchors the length to zero at the origin. In simulation, adding both losses gives the lowest held-out error whenever $N_{\mathrm{train}}$ is small, and the model with only PINN already improves extrapolation near joint-angle extremes where data are sparse; in the real robot, the physics loss must be weakened ($\alpha$ from $10^{-5}$ down to $10^{-7}$ or $10^{-8}$) to avoid the mismatch from friction, after which the combined method again matches or beats the conventional baseline.

Load-bearing premise

The method assumes the robot's joint structure is known accurately enough that the gravity-compensation torque at each joint angle can be treated as correct; when real friction violates that assumption the physics loss must be weakened or it actively degrades learning.

Editorial extensions

If this is right

Body-schema networks for tendon-driven robots can be trained from tens rather than hundreds of measured postures, shortening the data-collection phase.
The muscle Jacobian needed for state estimation and torque control comes out of the same trained network by differentiation, so the physics constraint doubles as a source of derivative supervision.
The benefit of the physics loss concentrates in the small-data regime and fades as $N_{\mathrm{train}}$ grows, so the method is most useful for new robots, new limbs, or re-learning after damage.
On physical hardware the physics weight $\alpha$ must be tuned down to offset friction and joint-model error; with that tuning the combined loss again matches or beats the plain-data baseline.
Introducing the origin constraint $L_{\mathrm{const}}$ consistently helps in both simulation and hardware, and is a zero-cost addition to any body-schema network.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The same torque-tension consistency loss could be applied to any learned actuator-to-joint map, not only muscle length, whenever a kinematic or dynamic model of the structure is available.
The real-robot result suggests replacing the fixed scalar $\alpha$ with an automatically calibrated or heteroscedastic weight that down-weights the physics term exactly where data disagree with it; the paper does not test this.
Because $L_{\mathrm{pinn}}$ provides gradient information in regions with no data, the method may generalize to other redundant, tendon-driven systems such as soft or cable-driven manipulators where geometric models are also unreliable.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Main take: this is a decent paper with an honest core idea — use the torque–tension relationship as a soft PINN loss when learning muscle length maps for musculoskeletal robots — and the simulation results are genuinely impressive at very small N_train. The real-robot part, though, is weaker than the abstract makes it sound.

What's new: applying PINN to body schema learning is not a big leap, but it's a sensible application and the paper doesn't oversell it. The derivation (Eq. 8 as a constraint on the learned Jacobian) is clean. The L_const origin constraint is simple and effective. In the 2-DOF simulation, Basic+Const+PINN cuts eval error by 60–70% at N_train=3–5 for AL-Map, and 20–35% for ATL-Map. Those are real gains with clear visual evidence of better extrapolation at joint limits. The paper also deserves credit for plotting training curves, reporting multiple seeds, and being upfront about friction and model mismatch — Section IV is an honest discussion.

Where it gets soft. First, the actual-robot comparison in Table III. The authors vary alpha over four values and then report the best evaluation value for each N_train, with no held-out validation set. That's selection on the test set, and it inflates the expected benefit. At N_train=10 the winning alpha gives 235 vs 257 for Basic+Const — an 8% drop with overlapping standard deviations. At N_train=30 the three best methods (18.17, 18.39, 18.99) all overlap within one standard deviation. So the real-robot evidence supports 'not worse' more than 'higher performance.' Second, the simulation is a consistency check: data and physics loss come from the same geometric model, so the strong simulation gains are what you'd expect. The paper acknowledges this implicitly by showing that on the real robot, the simulation-weight alpha (1e-5) blows up the error. Third, there's no code or data release, and no significance testing; with n=5 seeds, the claim of superiority on the real system needs an error bar on the difference. The alpha-dependence (best alpha 1e-8 at N=10, 1e-7 at N=30) reinforces that the right strength of the physics prior is data-dependent and needs a principled selection rule.

Who's it for: people working on musculoskeletal humanoids and anyone thinking about using PINN-style losses with imperfect physical priors. It deserves a serious referee: the idea is coherent, the simulation evidence is strong, and the real-robot weaknesses are fixable with a proper validation split and significance analysis. I'd recommend reviewing it, with a request to revise the real-robot claim to what the data actually support.

Referee Report

4 major / 5 minor

Summary. The paper proposes PIMBS, a physics-informed neural network approach for learning the body schema of musculoskeletal humanoids, specifically the mapping from joint angles (and, in the ATL-Map variant, muscle tension) to muscle length. The method augments the standard mean-squared-error loss with two additional terms: a physics loss L_pinn that enforces the static equilibrium relation τ = -G^T f through the network's Jacobian, and a constraint loss L_const that enforces zero muscle length at the zero configuration. The approach is evaluated on a 2-DOF 4-muscle simulation for both AL-Map and ATL-Map formulations and on the 5-DOF 10-muscle left arm of a real musculoskeletal humanoid, with training set sizes from 3 to 30. The simulation results show large and consistent improvements of Basic+Const+PINN over the conventional Basic baseline when the training set is small, while the real-robot results show smaller and noisier gains that depend strongly on the chosen physics-loss weight α.

Significance. The paper addresses a real bottleneck in musculoskeletal humanoid research: the high cost and difficulty of collecting large amounts of ground-truth body-schema data. If the claimed data-efficiency were firmly established, the method would be practically valuable, and the idea of leveraging the muscle Jacobian as a physics-informed regularizer is a sensible and nontrivial contribution. The simulation experiments are a clear strength: Tables I and II show that Basic+Const+PINN reduces held-out muscle-length error by roughly 60-70% (AL-Map) and 20-35% (ATL-Map) compared to Basic at Ntrain=3-5 and 5-10, respectively, with modest standard deviations. The physics loss is a genuine external constraint derived from mechanics rather than from the target outputs, and the use of a real musculoskeletal humanoid in Section III-D is a valuable independent test. However, the real-robot evidence currently does not support the paper's central claim of superior performance on the actual robot, because the comparison selects the physics weight α using the evaluation set and because the best-case differences at Ntrain=30 are within one standard deviation.

major comments (4)

[Section III-D, Table III] The real-robot comparison is distorted by selection on the evaluation set. The text states that α ∈ {1e-5, 1e-6, 1e-7, 1e-8} is varied, and Table III reports the Lbest_eval values for each α. No held-out validation set is used to choose α or the early-stopping epoch; the best value is effectively selected in hindsight on Deval. This is a load-bearing issue because the winning α differs by Ntrain (1e-8 for Ntrain=10, 1e-7 for Ntrain=30), which indicates that the reported gain is not a stable property of the method. Moreover, at Ntrain=30, all Basic+Const+PINN values (18.17±4.79, 18.39±4.51) overlap with Basic+Const (18.99±4.38) within one standard deviation, and at Ntrain=10, α=1e-7 (276.16±119.48) is worse than Basic+Const (256.57±72.48) while α=1e-8 (235.31±65.11) is only about 8% better. Without an untouched test set and a pre-specified α, the claim that PIMBS is more efficient on the real robot is not established.
[Sections II-C and III-B/III-C] In the simulation experiments, both the training data and the physics loss are derived from the same geometric model: the exponential spring law f = exp(K·Δn) - 1 in Eq. (13) is used to generate muscle lengths, and the gravity torque τ is computed geometrically with the same model. This makes L_pinn exactly consistent with the data generator by construction, so the simulated advantage partly reflects a consistency check rather than robustness to the model mismatch that motivates the real-robot study. The real-robot experiment is the independent test, and, as discussed above, that test is currently too weak to support the generalization of the simulation gains.
[Section III-D and Section IV] The paper acknowledges that friction and joint-model error invalidate Eq. (8) on the real robot, and that α must be reduced from 1e-5 to 1e-8 to avoid performance degradation. However, the paper provides no principled mechanism for setting α in practice; the only guidance is that smaller α works better empirically, and the best α is data-dependent. Since α is chosen using the evaluation metric, the real-robot results overstate the benefit of the method under a realistic deployment scenario where ground-truth muscle lengths are scarce. A sensitivity analysis with α fixed a priori, or a proper validation-split protocol for α selection, would be necessary to make the claim credible.
[Section V (Conclusion)] The conclusion states that the approach enables 'more efficient body schema learning from a small amount of data with higher performance than conventional methods in both simulation and the actual musculoskeletal humanoid.' Given the evaluation-set selection issue and the overlapping standard deviations in Table III, the real-robot part of this claim is not supported by the presented evidence. The conclusion should be qualified to reflect the simulation results as the primary support, with the real-robot results described as preliminary or as a demonstration of feasibility requiring further validation.

minor comments (5)

[Section III-D] The text contains a formatting error: 'α = 1.0 × 10{−5,−6,−7,−8}' should be rendered as α = 1.0 × 10^{-5,-6,-7,-8} with superscripts.
[Section III-A] The paper says 'all displayed loss values are scaled by a factor of 1.0 × 10^5 for clearer visualization,' but Tables I-III report values that are not obviously scaled and the text uses the same Lbest_eval notation for both scaled and unscaled values. Please clarify whether the tables show scaled or raw values.
[Section III-A] For the real-robot experiments, it is not specified how the five repeated runs (which produce the standard deviations in Table III) are constructed from the 496 collected data points: is the Ntrain subset randomly drawn for each seed, or are the same points used across seeds? This affects the interpretation of the reported variance.
[Section II-A / Eq. (8)] Eq. (8) is used as the physics loss for both AL-Map and ATL-Map, but for the ATL-Map the network output also depends on f, and the Jacobian in Eq. (9) is evaluated at the observed f. The text could state more explicitly that τ is the gravity-compensation torque computed from the geometric model rather than a directly measured quantity; this is important for understanding the sensor/model assumptions behind L_pinn.
[References] Reference [16] has a capitalization error ('Journal of Computational physics'), and the paper would benefit from a brief comparison with or citation of more recent PINN applications in robotics, if any exist.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the physics loss is an external mechanical constraint and the origin constraint is a definitional boundary condition.

full rationale

The derivation chain is self-contained, and no load-bearing step reduces to its own inputs. The core addition is the PINN loss Lpinn = mean((Gpred^T f_data + tau_data)^2), which couples the network Jacobian Gpred to the gravity-compensation torque tau through the mechanical equilibrium relation tau = -G^T f. This is an external constraint that does not use the target muscle-length labels, and on the real robot it is approximate enough to hurt performance unless the weight alpha is reduced, demonstrating that it carries independent content rather than being a restatement of the data. The Lconst term enforces l = 0 at theta = 0, f = 0, which is exactly the paper's declared convention that muscle length is expressed as a relative change with value at theta = 0 set to 0; it is a known boundary condition, not a re-derivation of the predicted output from that output. In simulation, the training data and the physics torque are generated from the same geometric model, but this makes the physics loss a consistency constraint rather than a circular one, and the real-robot experiment provides an independent test. Self-citations such as [4] and [13]-[15] identify the robot platform and prior baselines but are not used as evidence for the efficiency claim. The practice of reporting the best alpha among {1e-5, ..., 1e-8} on the evaluation set is a statistical-selection concern, not a circularity of the derivation. Overall, the method's claimed improvement is supported by an external physical constraint and an explicit boundary condition, with no fitted parameter renamed as a prediction.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The method relies on the correctness of the joint model, static torque balance, and slack-free data. The physics loss weight alpha is a tuned hyperparameter, and the simulation uses a specific spring model. No new physical entities are introduced.

free parameters (2)

alpha (weight of physics loss Lpinn) = 1e-5 in simulation; swept over 1e-5 to 1e-8 in real robot
Hyperparameter controlling the influence of the physics loss. In real-robot experiments it is swept and the best value per data size is reported, which is a form of tuning on the evaluation set.
K in nonlinear elastic element model f = e^(K*Delta_n) - 1 = 1000
Chosen simulation constant that determines the stretch-tension relationship in the generated data. It is not fitted to a physical target and may not match real nonlinear elements, affecting external validity of simulation results.

assumptions (4)

domain assumption The robot's joint structure is correct enough to compute gravity compensation torque tau(theta).
Invoked in Eq. (8) and Section II-C. The paper acknowledges real-world violations due to friction and model error, and discusses this as future work in Section IV.
domain assumption Static force balance tau = -G^T f holds instantaneously for the collected data.
Underlies Lpinn. Ignores friction, dynamics, and hysteresis, which the discussion identifies as the main source of performance loss on the real robot.
domain assumption Muscle slack is eliminated during data collection.
Stated in Section II-B as a requirement for the measured muscle length to correspond to the model. If slack remains, the data-label relationship breaks down.
ad hoc to paper Simulation data uses the exponential spring model f = e^(K*Delta_n) - 1 with K=1000.
This model is chosen for simulation only and is not validated against the real robot. The simulation conclusions depend on this choice.

how reviews work

0 comments

Cite this review

Pith. "Pith review of PIMBS: Efficient Body Schema Learning for Musculoskeletal Humanoids with Physics-Informed Neural Networks." pith.science (2026). https://pith.science/paper/SLH7WS3J

@misc{pith2026250620343,
  author       = {Pith},
  title        = {Pith review of: PIMBS: Efficient Body Schema Learning for Musculoskeletal Humanoids with Physics-Informed Neural Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SLH7WS3J}},
  note         = {Machine review of arXiv:2506.20343}
}

read the original abstract

Musculoskeletal humanoids are robots that closely mimic the human musculoskeletal system, offering various advantages such as variable stiffness control, redundancy, and flexibility. However, their body structure is complex, and muscle paths often significantly deviate from geometric models. To address this, numerous studies have been conducted to learn body schema, particularly the relationships among joint angles, muscle tension, and muscle length. These studies typically rely solely on data collected from the actual robot, but this data collection process is labor-intensive, and learning becomes difficult when the amount of data is limited. Therefore, in this study, we propose a method that applies the concept of Physics-Informed Neural Networks (PINNs) to the learning of body schema in musculoskeletal humanoids, enabling high-accuracy learning even with a small amount of data. By utilizing not only data obtained from the actual robot but also the physical laws governing the relationship between torque and muscle tension under the assumption of correct joint structure, more efficient learning becomes possible. We apply the proposed method to both simulation and an actual musculoskeletal humanoid and discuss its effectiveness and characteristics.

Figures

Figures reproduced from arXiv: 2506.20343 by the authors.

**Figure 2.** The basic structure of musculoskeletal humanoids: muscle motors, [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

**Figure 4.** The training scheme of Physics-Informed Musculoskeletal Body [PITH_FULL_IMAGE:figures/full_fig_p003_4.png] view at source ↗

Figures from the paper (7 more)

**Figure 6.** Figure 6: The transitions of Leval when training AL-Map with 3 data points in the 2-DOF 4-muscle musculoskeletal simulation. Basic Basic+Const Basic+PINN Basic+Const+PINN AL-Map (𝑁 𝑡𝑟𝑎𝑖𝑛 = 5) [PITH_FULL_IMAGE:figures/full_fig_p004_6.png]

**Figure 7.** Figure 7: The transitions of Leval when training AL-Map with 5 data points in the 2-DOF 4-muscle musculoskeletal simulation. varying Ntrain = {3, 5, 10}. The results for Ntrain = 3, Ntrain = 5, and Ntrain = 10 are shown in [PITH_FULL_IMAGE:figures/full_fig_p004_7.png]

**Figure 8.** Figure 8: The transitions of Leval when training AL-Map with 10 data points in the 2-DOF 4-muscle musculoskeletal simulation. impact on performance than Const [PITH_FULL_IMAGE:figures/full_fig_p005_8.png]

**Figure 9.** Figure 9: The relationship between joint angle and muscle length when training AL-Map with 5 data points in the 2-DOF 4-muscle musculoskeletal [PITH_FULL_IMAGE:figures/full_fig_p006_9.png]

**Figure 10.** Figure 10: The transitions of Leval when training ATL-Map with 5 data points in the 2-DOF 4-muscle musculoskeletal simulation. Basic Basic+Const Basic+PINN Basic+Const+PINN ATL-Map (𝑁 𝑡𝑟𝑎𝑖𝑛 = 10) [PITH_FULL_IMAGE:figures/full_fig_p007_10.png]

**Figure 11.** Figure 11: The transitions of Leval when training ATL-Map with 10 data points in the 2-DOF 4-muscle musculoskeletal simulation. friction – difficult to model as static physical laws – introduce inconsistencies between the collected data and physical laws. Therefore, when physica…

**Figure 14.** Figure 14: The transitions of Leval when training ATL-Map with 30 data points in the left arm of the musculoskeletal humanoid. [9] H. G. Marques, F. Imtiaz, F. Iida, and R. Pfeifer, “Self-organization of reflexive behavior from spontaneous motor activity,” Biological Cybernetics…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

17 extracted references · 16 canonical work pages

[1]

ECCE1: the first of a series of anthropomimetic musculoskeletal upper torsos,

H. G. Marques, M. J ¨antsh, S. Wittmeier, O. Holland, C. Alessandro, A. Diamond, M. Lungarella, and R. Knight, “ECCE1: the first of a series of anthropomimetic musculoskeletal upper torsos,” in IEEE-RAS International Conference on Humanoid Robots , 2010, pp. 391–396

work page 2010
[2]

Anthrob - A Printed Anthropomimetic Robot,

M. J ¨antsch, S. Wittmeier, K. Dalamagkidis, A. Panos, F. V olkart, and A. Knoll, “Anthrob - A Printed Anthropomimetic Robot,” in IEEE- RAS International Conference on Humanoid Robots , 2013, pp. 342– 347

work page 2013
[3]

Human Mimetic Musculoskele- tal Humanoid Kengoro toward Real World Physically Interactive Ac- tions,

Y . Asano, T. Kozuki, S. Ookubo, M. Kawamura, S. Nakashima, T. Katayama, Y . Iori, H. Toshinori, K. Kawaharazuka, S. Makino, Y . Kakiuchi, K. Okada, and M. Inaba, “Human Mimetic Musculoskele- tal Humanoid Kengoro toward Real World Physically Interactive Ac- tions,” in IEEE-RAS International Conference on Humanoid Robots , 2016, pp. 876–883

work page 2016
[4]

Component Modularized Design of Musculoskeletal Hu- manoid Platform Musashi to Investigate Learning Control Systems,

K. Kawaharazuka, S. Makino, K. Tsuzuki, M. Onitsuka, Y . Nagamatsu, K. Shinjo, T. Makabe, Y . Asano, K. Okada, K. Kawasaki, and M. Inaba, “Component Modularized Design of Musculoskeletal Hu- manoid Platform Musashi to Investigate Learning Control Systems,” in IEEE/RSJ International Conference on Intelligent Robots and Systems , 2019, pp. 7294–7301

work page 2019
[5]

Control strategies for tendon-driven manipulators,

S. C. Jacobsen, H. Ko, E. K. Iversen, and C. C. Davis, “Control strategies for tendon-driven manipulators,” IEEE Control Systems Magazine, vol. 10, no. 2, pp. 23–28, 1990

work page 1990
[6]

Mechanical stiffness control of tendon-driven joints,

K. Koganezawa and M. Yamazaki, “Mechanical stiffness control of tendon-driven joints,” in IEEE/RSJ International Conference on Intelligent Robots and Systems , 1999, pp. 818–825

work page 1999
[7]

Robust Continuous Motion Strategy Against Muscle Rupture using Online Learning of Redundant Intersensory Networks for Musculoskeletal Humanoids,

K. Kawaharazuka, M. Nishiura, Y . Toshimitsu, Y . Omura, Y . Koga, Y . Asano, K. Okada, K. Kawasaki, and M. Inaba, “Robust Continuous Motion Strategy Against Muscle Rupture using Online Learning of Redundant Intersensory Networks for Musculoskeletal Humanoids,” Robotics and Autonomous Systems , vol. 152, pp. 1–14, 2022

work page 2022
[8]

Reflex-like control for the coupled tendon- driven manipulator,

H. Endo and M. Wada, “Reflex-like control for the coupled tendon- driven manipulator,” in IEEE/RSJ International Conference on Intel- ligent Robots and Systems , vol. 3, 1994, pp. 1810–1815. Basic Basic+Const Basic+PINN Basic+Const+PINN ATL-Map (𝑁𝑡𝑟𝑎𝑖𝑛 = 30) Basic+Const+PINN (𝛼 = 10−5) Basic+Const+PINN (𝛼 = 10−6) Basic+Const+PINN (𝛼 = 10−7) Basic+Const+PI...

work page 1994

Show all 17 references

[9]

Self-organization of reflexive behavior from spontaneous motor activity,

H. G. Marques, F. Imtiaz, F. Iida, and R. Pfeifer, “Self-organization of reflexive behavior from spontaneous motor activity,” Biological Cybernetics, vol. 107, no. 1, pp. 25–37, 2013

2013
[10]

Robotic investigation on effect of stretch reflex and crossed inhibitory response on bipedal hopping,

X. Liu, A. Rosendo, S. Ikemoto, M. Shimizu, and K. Hosoda, “Robotic investigation on effect of stretch reflex and crossed inhibitory response on bipedal hopping,” Journal of The Royal Society Interface , vol. 15, no. 140, p. 20180024, 2018

2018
[11]

Joint proprio- ception acquisition strategy based on joints-muscles topological maps for musculoskeletal humanoids,

Y . Nakanishi, K. Hongo, I. Mizuuchi, and M. Inaba, “Joint proprio- ception acquisition strategy based on joints-muscles topological maps for musculoskeletal humanoids,” in IEEE International Conference on Robotics and Automation , 2010, pp. 1727–1732

2010
[12]

Learning Nonlinear Muscle-Joint State Mapping Toward Geometric Model-Free Tendon Driven Musculoskeletal Robots,

S. Ookubo, Y . Asano, T. Kozuki, T. Shirai, K. Okada, and M. Inaba, “Learning Nonlinear Muscle-Joint State Mapping Toward Geometric Model-Free Tendon Driven Musculoskeletal Robots,” in IEEE-RAS International Conference on Humanoid Robots , 2015, pp. 765–770

2015
[13]

Online Learning of Joint-Muscle Mapping using Vision in Tendon-driven Musculoskeletal Humanoids,

K. Kawaharazuka, S. Makino, M. Kawamura, Y . Asano, K. Okada, and M. Inaba, “Online Learning of Joint-Muscle Mapping using Vision in Tendon-driven Musculoskeletal Humanoids,” IEEE Robotics and Automation Letters, vol. 3, no. 2, pp. 772–779, 2018

2018
[14]

Long-time Self-body Image Acquisition and its Application to the Control of Musculoskeletal Structures,

K. Kawaharazuka, K. Tsuzuki, S. Makino, M. Onitsuka, Y . Asano, K. Okada, K. Kawasaki, and M. Inaba, “Long-time Self-body Image Acquisition and its Application to the Control of Musculoskeletal Structures,” IEEE Robotics and Automation Letters , vol. 4, no. 3, pp. 2965–2972, 2019

2019
[15]

Musculoskeletal AutoEncoder: A Uni- fied Online Acquisition Method of Intersensory Networks for State Estimation, Control, and Simulation of Musculoskeletal Humanoids,

K. Kawaharazuka, K. Tsuzuki, M. Onitsuka, Y . Asano, K. Okada, K. Kawasaki, and M. Inaba, “Musculoskeletal AutoEncoder: A Uni- fied Online Acquisition Method of Intersensory Networks for State Estimation, Control, and Simulation of Musculoskeletal Humanoids,” IEEE Robotics and...

2020
[16]

Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations,

M. Raissi, P. Perdikaris, and G. E. Karniadakis, “Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations,” Journal of Computational physics , vol. 378, pp. 686–707, 2019

2019
[17]

Adam: A Method for Stochastic Optimiza- tion,

D. P. Kingma and J. Ba, “Adam: A Method for Stochastic Optimiza- tion,” in 3rd International Conference on Learning Representations , 2015, pp. 1–15

2015

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

ECCE1: the first of a series of anthropomimetic musculoskeletal upper torsos,

H. G. Marques, M. J ¨antsh, S. Wittmeier, O. Holland, C. Alessandro, A. Diamond, M. Lungarella, and R. Knight, “ECCE1: the first of a series of anthropomimetic musculoskeletal upper torsos,” in IEEE-RAS International Conference on Humanoid Robots , 2010, pp. 391–396

work page 2010

[2] [2]

Anthrob - A Printed Anthropomimetic Robot,

M. J ¨antsch, S. Wittmeier, K. Dalamagkidis, A. Panos, F. V olkart, and A. Knoll, “Anthrob - A Printed Anthropomimetic Robot,” in IEEE- RAS International Conference on Humanoid Robots , 2013, pp. 342– 347

work page 2013

[3] [3]

Human Mimetic Musculoskele- tal Humanoid Kengoro toward Real World Physically Interactive Ac- tions,

Y . Asano, T. Kozuki, S. Ookubo, M. Kawamura, S. Nakashima, T. Katayama, Y . Iori, H. Toshinori, K. Kawaharazuka, S. Makino, Y . Kakiuchi, K. Okada, and M. Inaba, “Human Mimetic Musculoskele- tal Humanoid Kengoro toward Real World Physically Interactive Ac- tions,” in IEEE-RAS International Conference on Humanoid Robots , 2016, pp. 876–883

work page 2016

[4] [4]

Component Modularized Design of Musculoskeletal Hu- manoid Platform Musashi to Investigate Learning Control Systems,

K. Kawaharazuka, S. Makino, K. Tsuzuki, M. Onitsuka, Y . Nagamatsu, K. Shinjo, T. Makabe, Y . Asano, K. Okada, K. Kawasaki, and M. Inaba, “Component Modularized Design of Musculoskeletal Hu- manoid Platform Musashi to Investigate Learning Control Systems,” in IEEE/RSJ International Conference on Intelligent Robots and Systems , 2019, pp. 7294–7301

work page 2019

[5] [5]

Control strategies for tendon-driven manipulators,

S. C. Jacobsen, H. Ko, E. K. Iversen, and C. C. Davis, “Control strategies for tendon-driven manipulators,” IEEE Control Systems Magazine, vol. 10, no. 2, pp. 23–28, 1990

work page 1990

[6] [6]

Mechanical stiffness control of tendon-driven joints,

K. Koganezawa and M. Yamazaki, “Mechanical stiffness control of tendon-driven joints,” in IEEE/RSJ International Conference on Intelligent Robots and Systems , 1999, pp. 818–825

work page 1999

[7] [7]

Robust Continuous Motion Strategy Against Muscle Rupture using Online Learning of Redundant Intersensory Networks for Musculoskeletal Humanoids,

K. Kawaharazuka, M. Nishiura, Y . Toshimitsu, Y . Omura, Y . Koga, Y . Asano, K. Okada, K. Kawasaki, and M. Inaba, “Robust Continuous Motion Strategy Against Muscle Rupture using Online Learning of Redundant Intersensory Networks for Musculoskeletal Humanoids,” Robotics and Autonomous Systems , vol. 152, pp. 1–14, 2022

work page 2022

[8] [8]

Reflex-like control for the coupled tendon- driven manipulator,

H. Endo and M. Wada, “Reflex-like control for the coupled tendon- driven manipulator,” in IEEE/RSJ International Conference on Intel- ligent Robots and Systems , vol. 3, 1994, pp. 1810–1815. Basic Basic+Const Basic+PINN Basic+Const+PINN ATL-Map (𝑁𝑡𝑟𝑎𝑖𝑛 = 30) Basic+Const+PINN (𝛼 = 10−5) Basic+Const+PINN (𝛼 = 10−6) Basic+Const+PINN (𝛼 = 10−7) Basic+Const+PI...

work page 1994