Pith. sign in

REVIEW 4 major objections 6 minor 39 references

Hardware Implementation of a Zero-Prior-Knowledge Approach to Lifelong Learning in Kinematic Control of Tendon-Driven Quadrupeds

T0 review · 4 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read A tendon-driven quadruped learns a non-convex cyclic trajectory on hardware with zero prior body model, after five minutes of motor babbling and fifteen short refinements.

desk verdict A real hardware demo of the G2P pipeline on a four-limb tendon-driven quadruped, but the headline claim about learning 'non-convex non-differentiable' movements is contradicted by the paper's own task definition. read the letter →

arxiv 2508.15160 v1 pith:W3DEZI2P submitted 2025-08-21 cs.RO

Hesam Azadjou , Suraj Chakravarthi Raja , Ali Marjaninejad , Francisco J. Valero-Cuevas This is my paper · ORCID
classification cs.RO
keywords tendon-drivenrobotsmotorbabblingG2Palgorithmlifelonglearninginversekinematicshardware-in-the-loopbio-inspiredcontrolquadruped
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

General-to-Particular (G2P) is a two-stage learning rule: a robot first moves its joints randomly while collecting paired kinematics and motor activations, then uses the resulting inverse map as a warm start for short task-specific refinement trials. The paper applies G2P to a four-legged tendon-driven robot whose legs each have two joints and three tendons, and reports that the robot can follow a non-convex, non-differentiable cyclic trajectory in air after about five minutes of babbling and fifteen 20-second refinements. Joint-angle RMSE drops from 65.4° to 9.3°. The point of the paper is that a hardware system with redundant, over- and under-actuated tendons can acquire a usable control law without any explicit body model, dynamics equation, or precomputed inverse kinematics. If true, the result argues that lifelong motor learning can be achieved by exploration plus cumulative re-tuning rather than by system identification.

What carries the argument

The central mechanism is the General-to-Particular (G2P) two-phase learning pipeline. Phase one is generalized motor babbling: random tendon activations drive the limb while joint angles, velocities, and accelerations are recorded, and this paired sensory-motor data trains an ANN inverse map from 6D kinematics to 3D activations. Phase two is particular refinement: the desired task trajectory is fed through the ANN to generate activations, the resulting kinematics are appended to the training set, and the ANN is warm-started and re-trained; a corrective position-error feedback term is added to the command. The same pipeline is used per limb with four identical networks, one for each leg.

What would settle it

Compute the coverage of the babbling dataset around the target orbit, then re-run the fifteen refinements with babbling shortened to 60 seconds. If the RMSE curve is unchanged, the five-minute babbling claim is not load-bearing; if it degrades sharply, the coverage premise is confirmed as necessary.

Watch

Extended reading notes

Core claim

The paper claims that a physical tendon-driven quadruped can learn to reproduce a target cyclical movement in air using only random motor babbling followed by incremental refinements, with zero prior knowledge of its body or environment. Each limb's two joint angles are sensed and mapped by a single-hidden-layer multilayer perceptron to the three tendon activations that produce them; babbling data trains an initial inverse map, and each refinement re-trains that same network on a cumulative dataset. After fifteen refinements, the robot tracks the desired joint-angle trajectories with an RMSE reduced from 65.4° to 9.3°, and its endpoint paths approximate the target non-convex orbit. The autho

Load-bearing premise

The load-bearing premise is that a few minutes of random motor babbling leaves enough samples near the target cyclic trajectory for a one-hidden-layer linear MLP to learn an inverse map that transfers; the paper does not verify state-space coverage, and it states two different babbling durations (60 seconds in Methods II.D.1, five minutes in Results III.A).

Editorial extensions

If this is right

  • A tendon-driven robot with two joints and three tendons per limb can be controlled without computing inverse dynamics or explicit model parameters.
  • A single babbling session can serve as the common prior for several later tasks, because refinements re-tune the same network on cumulative data.
  • Because the network is warm-started and re-trained after every trial, the approach supports gradual adaptation to body changes such as wear-and-tear or altered actuator strength.
  • Adding a third physical dimension to the limbs would expand the activation space and is expected to require longer exploration, as the paper's own limitation section states.

Reading between the lines

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

  • The paper presents one demonstration run without repeated trials, so the reported RMSE improvement is a proof-of-concept number, not yet a measured performance distribution.
  • The abstract and conclusion say five minutes of babbling while Methods Section II.D.1 says 60 seconds; unless resolved, the time-to-learn claim is ambiguous and needs replication with a stated duration.
  • A sharper test of the General-to-Particular premise would be to push the target trajectory outside the region visited during babbling; the current experiment does not show whether the initial inverse map generalizes or the refinements do the work.
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.

Referee Report

4 major / 6 minor

Summary. The paper reports a hardware implementation of the General-to-Particular (G2P) algorithm on a custom, 3D-printed tendon-driven quadruped. Each leg is a planar two-DoF mechanism actuated by three tendons. The system first performs an exploratory motor-babbling phase, then uses a single-hidden-layer MLP as an inverse kinematics-to-activation map, and iteratively refines this map on task-specific trials. The authors claim that, with zero prior knowledge of the body or environment, the quadruped learns to follow a cyclical, non-convex, non-differentiable joint-angle trajectory in air: after five minutes of babbling and 15 refinement trials of 20 seconds each, the joint-angle RMSE is reduced from 65.4° to 9.3°. The paper includes the hardware design, the ANN training pipeline, the refinement procedure, and qualitative/quantitative results (Figs. 3-5).

Significance. If the central claims are correct, the work would be a useful proof-of-concept for model-free, bio-inspired learning in a genuinely redundant tendon-driven system, and would strengthen the case that few-shot, hardware-in-the-loop refinement can replace explicit dynamic models. The paper has clear strengths: a real four-legged robot built in-house, a detailed description of the electromechanical setup, a plausible online-learning loop, and a quantitative improvement that is visually supported by the figures. However, the evidence presented is narrower than the claims. The task description is internally contradictory, the babbling duration is reported inconsistently, and the central quantitative result comes from a single run with no baseline or repeated-run statistics. The 'lifelong learning' and 'non-convex non-differentiable' claims are not substantiated by the experiments as currently described. These issues do not necessarily invalidate the core idea, but they must be resolved before the claims can be accepted.

major comments (4)
  1. [II.B.1 / III.B / Abstract / Conclusion] The headline claim is that the robot learns 'non-convex non-differentiable cyclical movements' (Abstract, Conclusion), but Methods II.B.1 defines the task as 'both proximal and distal joints follow sinusoidal trajectories with a π/2 phase difference.' A phase-shifted sine/cosine pair traces an ellipse in joint space, which is convex and differentiable. Section III.B instead defines the orbit by a 10-point 'feature vector,' and Fig. 3 shows a clearly non-convex target path. The cycle duration also differs: II.B.1 says 3.6 s/cycle, while III.B says approximately 1 s/cycle. The authors must state which task was actually executed and reconcile these descriptions. As written, the abstract's central contribution is unsupported by the stated Methods.
  2. [II.D.1 / III.A / II.E] The duration of motor babbling is reported inconsistently. Section II.D.1 states 'we performed the babbling phase for 60 seconds' and Section II.E repeats 'a 60-second babbling'; but Section III.A and the Abstract say 'five-minute motor babbling,' and the Conclusion says 'just ten minutes (five minutes each).' This is a load-bearing discrepancy because the claimed time-to-learn is one of the paper's key selling points. Please correct the duration and recompute the total learning time accordingly.
  3. [III.C / Fig. 4] The main quantitative result—RMSE reduction from 65.4° to 9.3°—is reported for a single continuous run. Figure 4 shows only one sequence of boxplots (eight joint values per refinement), with no repeated runs, confidence intervals, or statistical test. There is also no baseline comparison: for example, the same task without refinements, or a standard inverse-kinematics/dynamics controller. Without repeated runs, the reported improvement cannot be distinguished from run-specific variation or from warm-up effects in a single long trial. The authors should report at least a few independent runs and include a no-refinement or alternative-control condition.
  4. [III.C / Conclusion / Title] The title and abstract claim 'lifelong learning' and adaptation to changing body/environment conditions, but the experiment involves a single fixed cyclical task in air, with no perturbation, task switch, or environmental change. Section II.E explicitly acknowledges that loading and environmental dynamics were not tested, yet the paper still presents lifelong-adaptation claims. Either demonstrate adaptation to at least one disturbance or parameter change, or temper the claims to 'few-shot refinement for a fixed task' in the title and abstract.
minor comments (6)
  1. [II.C, Eq. (1)] The second-order model in Eq. (1) appears malformed: the terms involving C(q, dot-q) and B dot-q have inconsistent matrix/vector grouping, and the sign/structure of the damping term is unclear. Please rewrite with explicit parentheses and verify dimensional consistency.
  2. [II.C] The text says the system uses 'four DC motors,' but Fig. 1 shows three motors per leg, so the quadruped should have twelve motors total. Please correct the number.
  3. [II.E / II.D.1] Section II.D.1 mentions tactile sensory inputs such as endpoint force values, but the hardware description mentions only encoders. Please clarify whether force/tactile data are actually used in this implementation, since the input vector described in Eq. (3) is kinematic only.
  4. [II.F] The sentence '24 neurons for the single ANN and 6 neurons for each ANN in the multiple ANN case' is unclear because 'single ANN' vs 'multiple ANN case' is not defined. Earlier the text says one ANN per limb (four total); please clarify the architecture and the distinction.
  5. [II.G] RMSE is computed over 'the last half of the data' to avoid transient effects. This is not precisely defined for trials of variable length. Please specify the exact evaluation window (e.g., last N cycles after the first transient cycle).
  6. [III.C / Fig. 4 caption] The caption says '15 refinements of a single over five minutes of continuous operation'—likely missing the word 'run.' Please rephrase for clarity.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the G2P algorithm is prior published work implemented in hardware, and the reported RMSE is a training-domain performance measure, not a forced prediction.

full rationale

The paper's derivation chain does not reduce to its own inputs by construction. The G2P algorithm is imported from prior work [17] by overlapping authors, but that prior result is a published, externally reviewed algorithm, and the present paper's contribution is an independent hardware implementation and evaluation; the self-citations are not used to forbid alternatives or to import an unverified uniqueness claim. The motor babbling phase provides independent exploratory sensorimotor data, and the refinement phase adds closed-loop task-specific data. The final RMSE reduction from 65.4° to 9.3° is measured on the desired cyclical trajectory that was also used during refinement, so it is a within-task performance measure rather than evidence of generalization to a new task. However, the paper does not present this RMSE as a derived law or as a prediction from first principles, and the ANN's supervised training targets are the actual activations that produced observed kinematics, not the desired trajectory itself. Thus no fitted parameter is silently renamed as a prediction. I also note two non-circular internal-consistency concerns that do not constitute circularity: Methods II.D.1 says the babbling phase lasted 60 seconds while Results III.A says five minutes, and the target is described in II.B.1 as sinusoidal (differentiable) yet the conclusions repeatedly call it non-differentiable and non-convex; these are correctness/reporting issues, not self-referential reductions. No quoted equation or fitted quantity can be exhibited as equivalent to the paper's stated input by construction, so the circularity score is 0.

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

The central claim rests on the expressive power of the MLP, the representativeness of babbling data, hardware stability, and task feasibility. No new physical entities are introduced. The paper provides no formal verification or independent benchmark for the inverse-map learning.

free parameters (4)
  • Input scaling factors = variance-derived from a 60-second babbling run
    Inputs are divided by their expected variance, computed from a single babbling run (Section II.E). This data-derived normalization affects all subsequent training.
  • ANN architecture = 1 hidden layer, 24 neurons for single ANN, 6 for multiple ANN, linear activation
    The architecture is a design choice not derived from theory (Section II.F). It is not fitted to the task, but it is a free choice affecting results.
  • Feature vector length for trajectory parameterization = 10 evenly distributed points
    The desired cyclical movement is parameterized by 10 points (Section III.B). This discretization is a choice that could affect the complexity of the learned mapping.
  • Number of refinements = 15 trials of 20 seconds
    The paper stops at 15 refinements; this is an experimental stopping point, not a rule. The outcome might change with more or fewer trials.
assumptions (4)
  • domain assumption The inverse map from limb kinematics to tendon activations is a smooth, stationary function that can be approximated by a single-hidden-layer MLP with linear activations.
    The paper assumes the ANN class is expressive enough to represent the inverse dynamics of the tendon-driven limb. This is not proven and could fail for highly nonlinear or time-varying dynamics.
  • domain assumption The motor babbling phase sufficiently explores the reachable state space of the limb so that the trained ANN generalizes to the target cyclic trajectory.
    Training data comes from random activations; the target is a specific sinusoidal pattern. The paper does not verify coverage of the state space or that the inverse map interpolates correctly (Section II.D).
  • domain assumption The hardware remains stable throughout the experiment, with no significant changes in tendon tension, motor temperature, or material properties.
    The paper mentions PLA warping under motor temperature in Section II.A but does not demonstrate stability during the runs. If hardware drifts, the learned inverse map may not remain valid.
  • domain assumption The desired cyclical movements in air are achievable given the actuation limits and mechanical constraints.
    The paper does not check feasibility of the target trajectories; it assumes the robot can physically produce them, and hard stops are set to avoid singularities (Section II.A).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Hardware Implementation of a Zero-Prior-Knowledge Approach to Lifelong Learning in Kinematic Control of Tendon-Driven Quadrupeds." pith.science (2026). https://pith.science/paper/W3DEZI2P

@misc{pith2026250815160,
  author       = {Pith},
  title        = {Pith review of: Hardware Implementation of a Zero-Prior-Knowledge Approach to Lifelong Learning in Kinematic Control of Tendon-Driven Quadrupeds},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/W3DEZI2P}},
  note         = {Machine review of arXiv:2508.15160}
}
read the original abstract

Like mammals, robots must rapidly learn to control their bodies and interact with their environment despite incomplete knowledge of their body structure and surroundings. They must also adapt to continuous changes in both. This work presents a bio-inspired learning algorithm, General-to-Particular (G2P), applied to a tendon-driven quadruped robotic system developed and fabricated in-house. Our quadruped robot undergoes an initial five-minute phase of generalized motor babbling, followed by 15 refinement trials (each lasting 20 seconds) to achieve specific cyclical movements. This process mirrors the exploration-exploitation paradigm observed in mammals. With each refinement, the robot progressively improves upon its initial "good enough" solution. Our results serve as a proof-of-concept, demonstrating the hardware-in-the-loop system's ability to learn the control of a tendon-driven quadruped with redundancies in just a few minutes to achieve functional and adaptive cyclical non-convex movements. By advancing autonomous control in robotic locomotion, our approach paves the way for robots capable of dynamically adjusting to new environments, ensuring sustained adaptability and performance.

Figures

Figures reproduced from arXiv: 2508.15160 by the authors.

Figure 1
Figure 1. Quadruped Leg Structure: Three tendons (label 1) are controlled by motors M0, M1, and M2. The mono-articular tendon connected to motor M1 exclusively drives the proximal joint in an counter-clockwise (flexion) direction. However the tendons connected to motors M0 and M2 are multi￾articular. Motor M0 drives the proximal joint clockwise (extension) and the distal joint counter-clockwise (flexion). Motor M2 only perfor… view at source ↗
Figure 2
Figure 2. System Overview: The physical implementation of the robot features software and hardware sections. Our modular tendon-driven robot, featuring DC motors and encoders, was controlled by a custom DAQ and control program on a DAQ-capable computer running the NI-DAQmx driver. The G2P learning pipeline—running on a machine learning-capable computer—used RT-Bridge over Ethernet to send commands and receive sensory data fro… view at source ↗
Figure 3
Figure 3. Endpoint trajectories of the quadruped Suspended in air: The green trajectories represent the performance immediately after babbling, with no refinement. The solid red trajectory indicates the target path the quadruped was supposed to follow. The blue trajectory illustrates the per￾formance after the final refinement. This figure demonstrates the quadruped’s progress following the target trajectory through successiv… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Refinements in Air: 15 refinements of a single over five minutes of continuous operation were performed in the air following motor babbling. The RMSE (Root Mean Square Error) of the resulting trajectories was plotted for the two joints of each of the four legs (eight v…
Figure 5
Figure 5. Figure 5: Joint Angles: The real joint angles were plotted vs the target trajectories after the 15th refinement showing the quality of the performance of the task by the legs for a cyclical motion in air (without load) with zero knowledge of the body and the environment before t…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

39 extracted references · 34 canonical work pages

  1. [1]

    Biological underpinnings for lifelong learning ma- chines,

    D. Kudithipudi, M. Aguilar-Simon, J. Babb, M. Bazhenov, D. Black- iston, J. Bongard, A. P. Brna, S. Chakravarthi Raja, N. Cheney, J. Clune, et al. , “Biological underpinnings for lifelong learning ma- chines,” Nature Machine Intelligence, vol. 4, no. 3, pp. 196–210, 2022

    work page 2022

  2. [2]

    Testing Skills in Vertebrates,

    M. S. Funk and P. Tosto, “Testing Skills in Vertebrates,” The American Biology Teacher , vol. 69, no. 2, pp. 85–91, Feb. 2007, publisher: National Association of Biology Teachers. [Online]. Available: https://bioone.org/journals/the-american-biology-teacher/ volume-69/issue-2/0002-7685 2007 69 85 TSIV 2.0.CO 2/ Testing-Skills-in-Vertebrates/10.1662/0002-7...

    work page doi:10.1662/0002-7685(2007)69 2007

  3. [3]

    Lifelong generative adversarial autoencoder,

    F. Ye and A. G. Bors, “Lifelong generative adversarial autoencoder,” IEEE Transactions on Neural Networks and Learning Systems , p. 1–15, 2023. [Online]. Available: http://dx.doi.org/10.1109/TNNLS. 2023.3281091

    arXiv 2023

  4. [4]

    Play it by ear: A perceptual algorithm for autonomous melodious piano playing with a bio-inspired robotic hand,

    H. Azadjou, A. Marjaninejad, and F. Valero-Cuevas, “Play it by ear: A perceptual algorithm for autonomous melodious piano playing with a bio-inspired robotic hand,” bioRxiv, pp. 2024–06, 2024

    work page 2024

  5. [5]

    De novo motor learning creates structure in neural activity space that shapes adaptation,

    J. C. Chang, M. G. Perich, L. E. Miller, J. A. Gallego, and C. Clopath, “De novo motor learning creates structure in neural activity space that shapes adaptation,” May 2023, pages: 2023.05.23.541925 Section: New Results. [Online]. Available: https://www.biorxiv.org/content/10.1101/2023.05.23.541925v2

    work page doi:10.1101/2023.05.23.541925v2 2023

  6. [6]

    Fine synergies

    A. Bartsch-Jimenez, M. Bła ˙zkiewicz, H. Azadjou, R. Novotny, and F. J. Valero-Cuevas, “Fine synergies” describe motor adaptation in people with drop foot in a way that supplements traditional “coarse synergies,” Frontiers in sports and active living , vol. 5, p. 1080170, 2023, publisher: Frontiers. [Online]. Available: https: //www.frontiersin.org/articl...

    work page arXiv 2023

  7. [7]

    Sim-to-Real via Sim-to-Sim: Data-efficient Robotic Grasping via Randomized-to-Canonical Adaptation Networks

    S. James, P. Wohlhart, M. Kalakrishnan, D. Kalashnikov, A. Irpan, J. Ibarz, S. Levine, R. Hadsell, and K. Bousmalis, “Sim-to-Real via Sim-to-Sim: Data-efficient Robotic Grasping via Randomized-to- Canonical Adaptation Networks,” July 2019, arXiv:1812.07252 [cs]. [Online]. Available: http://arxiv.org/abs/1812.07252

    work page Pith review arXiv 2019

  8. [8]

    Solving rubik’s cube with a robot hand,

    I. Akkaya, M. Andrychowicz, M. Chociej, M. Litwin, B. McGrew, A. Petron, A. Paino, M. Plappert, G. Powell, R. Ribas, et al., “Solving rubik’s cube with a robot hand,” arXiv preprint arXiv:1910.07113 , 2019

    arXiv 1910

Show all 39 references
  1. [9]

    Resilient Machines Through Continuous Self-Modeling,

    J. Bongard, V . Zykov, and H. Lipson, “Resilient Machines Through Continuous Self-Modeling,” Science, vol. 314, no. 5802, pp. 1118–1121, Nov. 2006, publisher: American Association for the Advancement of Science. [Online]. Available: https: //www.science.org/doi/10.1126/science.1133687

    2006 doi

  2. [10]

    Hierarchical reinforcement learning of multiple grasping strategies with human instructions,

    T. Osa, J. Peters, and G. Neumann, “Hierarchical reinforcement learning of multiple grasping strategies with human instructions,” Advanced Robotics , vol. 32, no. 18, pp. 955–968, Sept. 2018. [Online]. Available: https://www.tandfonline.com/doi/full/10.1080/ 01691864.2018.1509018

    2018

  3. [11]

    Reinforcement learning for non-prehensile manipulation: Transfer from simulation to physical system,

    K. Lowrey, S. Kolev, J. Dao, A. Rajeswaran, and E. Todorov, “Reinforcement learning for non-prehensile manipulation: Transfer from simulation to physical system,” in 2018 IEEE International Conference on Simulation, Modeling, and Programming for Autonomous Robots (SIMPAR) . Br...

    2018

  4. [12]

    Incremental skill acquisition for self-motivated learning animats,

    A. Bonarini, A. Lazaric, and M. Restelli, “Incremental skill acquisition for self-motivated learning animats,” in International Conference on Simulation of Adaptive Behavior . Springer, 2006, pp. 357–368

    2006

  5. [13]

    Self-organizing incremental neural net- work (soinn) as a mechanism for motor babbling and sensory-motor learning in developmental robotics,

    T. Najjar and O. Hasegawa, “Self-organizing incremental neural net- work (soinn) as a mechanism for motor babbling and sensory-motor learning in developmental robotics,” in International Work-Conference on Artificial Neural Networks . Springer, 2013, pp. 321–330

    2013

  6. [14]

    Brain- Body-Task Co-Adaptation can Improve Autonomous Learning and Speed of Bipedal Walking,

    D. Urbina-Mel ´endez, H. Azadjou, and F. J. Valero-Cuevas, “Brain- Body-Task Co-Adaptation can Improve Autonomous Learning and Speed of Bipedal Walking,” Feb. 2024, arXiv:2402.02387 [cs]. [Online]. Available: http://arxiv.org/abs/2402.02387

    2024 arXiv

  7. [15]

    Resilient machines through continuous self-modeling,

    J. Bongard, V . Zykov, and H. Lipson, “Resilient machines through continuous self-modeling,” Science, vol. 314, no. 5802, pp. 1118– 1121, 2006

    2006

  8. [16]

    Task-agnostic self-modeling ma- chines,

    R. Kwiatkowski and H. Lipson, “Task-agnostic self-modeling ma- chines,” Science Robotics, vol. 4, no. 26, p. eaau9354, 2019

    2019

  9. [17]

    Autonomous functional movements in a tendon-driven limb via limited experience,

    A. Marjaninejad, D. Urbina-Mel ´endez, B. A. Cohn, and F. J. Valero- Cuevas, “Autonomous functional movements in a tendon-driven limb via limited experience,” Nature machine intelligence, vol. 1, no. 3, p. 144, 2019

    2019

  10. [18]

    Should anthropomorphic systems be “redundant

    A. Marjaninejad and F. J. Valero-Cuevas, “Should anthropomorphic systems be “redundant”?” in Biomechanics of Anthropomorphic Sys- tems. Springer, 2019, pp. 7–34

    2019

  11. [19]

    F. J. Valero-Cuevas, Fundamentals of neuromechanics . Springer, 2016, vol. 8

    2016

  12. [20]

    Neuromorphic meets neuromechan- ics, part ii: the role of fusimotor drive,

    K. Jalaleddini, C. M. Niu, S. C. Raja, W. J. Sohn, G. E. Loeb, T. D. Sanger, and F. J. Valero-Cuevas, “Neuromorphic meets neuromechan- ics, part ii: the role of fusimotor drive,” Journal of neural engineering, vol. 14, no. 2, p. 025002, 2017

    2017

  13. [21]

    A Novel Passive Implantable Differential Mechanism to Restore Individuated Finger Flexion during Grasping following Tendon Transfer Surgery: A Pilot Study,

    S. Chakravarthi Raja, W. S. You, K. Jalaleddini, J. C. Casebier, N. R. Lightdale-Miric, V . R. Hentz, F. J. Valero-Cuevas, and R. Bal- asubramanian, “A Novel Passive Implantable Differential Mechanism to Restore Individuated Finger Flexion during Grasping following Tendon Tran...

    2023

  14. [22]

    [Online]

    AMT103-v. [Online]. Available: https://www.digikey.com/en/products/ detail/cui-devices/AMT103-V/827016

  15. [23]

    Marionette: A neurorobotics toolchain for the neuromorphic control of tendon-driven systems,

    S. Chakravarthi Raja, “Marionette: A neurorobotics toolchain for the neuromorphic control of tendon-driven systems,” Ph.D. dissertation, University of Southern California, Los Angeles, California, United States, Sept. 2023. [Online]. Available: https: //digitallibrary.usc.edu/...

    2023

  16. [24]

    Hintjens, ZeroMQ: messaging for many applications

    P. Hintjens, ZeroMQ: messaging for many applications . ” O’Reilly Media, Inc.”, 2013

    2013

  17. [25]

    How to design & implement a modern communication middleware based on zeromq,

    J. Lauener, W. Sliwinski, and G. CERN, “How to design & implement a modern communication middleware based on zeromq,” in Proc of ICALEPCS, vol. 17, 2017, pp. 45–51

    2017

  18. [26]

    Adam: A method for stochastic optimiza- tion,

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimiza- tion,” arXiv preprint arXiv:1412.6980 , 2014

    2014 arXiv

  19. [27]

    Sim- ple kinematic feedback enhances autonomous learning in bio-inspired tendon-driven systems,

    A. Marjaninejad, D. Urbina-Mel ´endez, and F. J. Valero-Cuevas, “Sim- ple kinematic feedback enhances autonomous learning in bio-inspired tendon-driven systems,” arXiv preprint arXiv:1907.04539 , 2019

    1907 arXiv

  20. [28]

    The effect of psychedelics on the level of brain-derived neurotrophic factor: A sys- tematic review and meta-analysis,

    A. Shafiee, R. Arabzadeh Bahri, M. A. Rafiei, F. Esmaeilpur Abi- aneh, P. Razmara, K. Jafarabady, and M. J. Amini, “The effect of psychedelics on the level of brain-derived neurotrophic factor: A sys- tematic review and meta-analysis,” Journal of Psychopharmacology , vol. 38, ...

    2024

  21. [29]

    Security analysis of a key based color image watermarking vs. a non-key based technique in telemedicine applications,

    P. N. Bideh, M. Mahdavi, S. E. Borujeni, and S. Arasteh, “Security analysis of a key based color image watermarking vs. a non-key based technique in telemedicine applications,” Multimedia Tools and Applications, vol. 77, pp. 31 713–31 735, 2018

    2018

  22. [30]

    Design and implementation of an ultralow-power ecg patch and smart cloud-based platform,

    B. Baraeinejad, M. F. Shayan, A. R. Vazifeh, D. Rashidi, M. S. Hamedani, H. Tavolinejad, P. Gorji, P. Razmara, K. Vaziri, D. Vashaee, et al., “Design and implementation of an ultralow-power ecg patch and smart cloud-based platform,” IEEE Transactions on Instrumentation and Mea...

    2022

  23. [31]

    Hybrid zero dynamics control of legged robots,

    A. D. Ames and I. Poulakakis, “Hybrid zero dynamics control of legged robots,” pp. 292–331, 2017

    2017

  24. [32]

    Zero-moment point—thirty five years of its life,

    M. Vukobratovi ´c and B. Borovac, “Zero-moment point—thirty five years of its life,” International journal of humanoid robotics , vol. 1, no. 01, pp. 157–173, 2004

    2004

  25. [33]

    The dynamics of legged locomotion: Models, analyses, and challenges,

    P. Holmes, R. J. Full, D. Koditschek, and J. Guckenheimer, “The dynamics of legged locomotion: Models, analyses, and challenges,” SIAM review, vol. 48, no. 2, pp. 207–304, 2006

    2006

  26. [34]

    The six determinants of gait and the inverted pendulum analogy: A dynamic walking perspective,

    A. D. Kuo, “The six determinants of gait and the inverted pendulum analogy: A dynamic walking perspective,” Human movement science , vol. 26, no. 4, pp. 617–656, 2007

    2007

  27. [35]

    Bifurcations and chaos in passive dynamic walking: A review,

    S. Iqbal, X. Zang, Y . Zhu, and J. Zhao, “Bifurcations and chaos in passive dynamic walking: A review,” Robotics and Autonomous Systems, vol. 62, no. 6, pp. 889–909, 2014

    2014

  28. [36]

    Stabilization of lateral motion in passive dynamic walking,

    A. D. Kuo, “Stabilization of lateral motion in passive dynamic walking,” The International journal of robotics research, vol. 18, no. 9, pp. 917–930, 1999

    1999

  29. [37]

    Dynamics and balance of a humanoid robot during manipulation tasks,

    K. Harada, S. Kajita, K. Kaneko, and H. Hirukawa, “Dynamics and balance of a humanoid robot during manipulation tasks,” IEEE Transactions on Robotics , vol. 22, no. 3, pp. 568–575, 2006

    2006

  30. [38]

    Spinal-like regulator facilitates control of a two-degree-of-freedom wrist,

    G. Raphael, G. A. Tsianos, and G. E. Loeb, “Spinal-like regulator facilitates control of a two-degree-of-freedom wrist,” Journal of Neu- roscience, vol. 30, no. 28, pp. 9431–9444, 2010

    2010

  31. [39]

    An α-mn collateral to γ-mns can mitigate velocity-dependent stretch reflexes during voluntary movement: A computational study,

    G. Niyo, L. I. Almofeez, A. Erwin, and F. J. Valero-Cuevas, “An α-mn collateral to γ-mns can mitigate velocity-dependent stretch reflexes during voluntary movement: A computational study,” bioRxiv, pp. 2023–12, 2023

    2023

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.