pith. sign in

arxiv: 2209.12133 · v3 · pith:GKUZYO55new · submitted 2022-09-25 · 💻 cs.RO · cs.SY· eess.SY

Development of a Deep Learning-Driven Control Framework for Exoskeleton Robots

Sk Hasan This is my paper

Pith reviewed 2026-05-24 10:42 UTC · model grok-4.3

classification 💻 cs.RO cs.SYeess.SY
keywords exoskeleton robotdeep neural networktorque controltrajectory trackinghybrid controllercomputational efficiencyseven DOFlower extremity
0
0 comments X

The pith

A hybrid deep neural network and PD controller achieves accurate trajectory tracking for seven-DOF exoskeleton robots while lowering computational demands compared to conventional nonlinear controllers.

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

This paper develops a deep learning-based control framework to overcome the real-time computational challenges of model-based controllers for high-degree-of-freedom exoskeleton robots. A four-layer neural network with 49 neurons is trained on data from the analytical dynamic model of a seven-DOF lower extremity exoskeleton to predict joint torques. In operation, the network outputs torque commands and a proportional-derivative controller corrects any residual errors, with stability proven analytically and robustness tested via statistical analysis. Simulations show the approach matches the performance of computed torque and sliding mode controllers but with reduced computational load.

Core claim

The paper establishes that a parallel-structured deep neural network trained on physics-based data from the robot's dynamic model can generate joint torque commands for precise trajectory tracking in a seven degree of freedom exoskeleton, with a PD compensator handling prediction residuals to yield performance comparable to standard nonlinear controllers at lower computational cost.

What carries the argument

The parallel structured deep neural network consisting of four layers with 49 densely connected neurons trained to predict joint torques from the analytical dynamic model.

If this is right

  • Stability of the hybrid control scheme is guaranteed by analytical proof.
  • The controller maintains robustness under parameter variations as shown by analysis of variance.
  • Trajectory tracking remains accurate with torque profiles similar to those from computed torque, sliding mode, and other conventional controllers.
  • Computational burden is reduced during real-time implementation.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the model-data training generalizes, similar networks could replace analytical models in other robotic control tasks where computation is a bottleneck.
  • Hardware validation on physical exoskeletons would be needed to confirm performance beyond simulation.
  • The framework might integrate with sensor feedback for adaptive control in varying human-robot interaction scenarios.

Load-bearing premise

The neural network trained solely on data from the analytical dynamic model will perform adequately on the actual physical exoskeleton without significant model mismatch or unmodeled dynamics affecting the predictions.

What would settle it

Experimental results from implementing the controller on a physical seven-DOF exoskeleton robot showing significantly larger tracking errors or instability compared to simulation would falsify the claim of practical applicability.

read the original abstract

The purpose of this study is to develop a computationally efficient deep learning based control framework for high degree of freedom exoskeleton robots to address the real time computational limitations associated with conventional model based control. A parallel structured deep neural network was designed for a seven degree of freedom human lower extremity exoskeleton robot. The network consists of four layers with 49 densely connected neurons and was trained using physics based data generated from the analytical dynamic model. During real time implementation, the trained neural network predicts joint torque commands required for trajectory tracking, while a proportional derivative controller compensates for residual prediction errors. Stability of the proposed control scheme was analytically established, and robustness to parameter variations was evaluated using analysis of variance. Comparative simulations were conducted against computed torque, model reference computed torque, sliding mode, adaptive, and linear quadratic controllers under identical robot dynamics. Results demonstrate accurate trajectory tracking with torque profiles comparable to conventional nonlinear controllers while reducing computational burden. These findings suggest that the proposed deep learning based hybrid controller offers an efficient and robust alternative for controlling multi degree of freedom exoskeleton robots.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The manuscript proposes a hybrid controller for a 7-DOF lower-extremity exoskeleton in which a four-layer densely connected neural network (49 neurons) is trained exclusively on torque data generated from the analytical dynamic model; the NN supplies feedforward torques for trajectory tracking while a PD compensator corrects residuals. Analytical stability is claimed for the closed-loop system, robustness to parameter variations is asserted via ANOVA, and comparative simulations against computed-torque, sliding-mode, adaptive, and LQR controllers are reported to show comparable tracking accuracy with reduced computational cost.

Significance. If the performance claims survive evaluation under realistic model mismatch, the work would supply a concrete, analytically grounded route to lowering the real-time computational load of high-DOF exoskeleton control while retaining stability guarantees. The provision of an analytical stability argument and side-by-side simulation comparisons against multiple established nonlinear controllers are positive features that would strengthen the contribution.

major comments (3)
  1. [Robustness evaluation / ANOVA results] The robustness section (ANOVA on parameter variations) evaluates performance only when the test dynamics remain identical to the analytical model used for NN training; no Monte-Carlo trials or sensitivity analysis under unmodeled effects (payload variation, joint friction, or inertial mismatch) are presented, leaving the claim that the PD compensator renders the scheme “robust” for physical hardware unsupported.
  2. [Comparative simulation results] All reported trajectory-tracking and torque-comparison results (including the computational-burden reduction) are generated under the exact same analytical dynamics for both training and testing; consequently the simulations do not probe the model-mismatch regime that the weakest assumption identifies as critical for the physical-robot claim.
  3. [Stability analysis] The analytical stability proof assumes bounded NN prediction error that the PD term can dominate, yet no explicit bound or Lipschitz analysis on the residual torque error is supplied when the inertia or Coriolis terms deviate from the training distribution; this gap directly affects the load-bearing stability guarantee.
minor comments (2)
  1. [Abstract] The abstract states that the network “consists of four layers with 49 densely connected neurons” but does not specify the activation functions, training algorithm, or loss metric; these details should be added for reproducibility.
  2. [Figures] Figure captions and axis labels for the torque and tracking-error plots should explicitly state the simulation duration, sampling rate, and whether the plotted trajectories are for a single representative trial or averaged over multiple runs.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback. We agree that the presented simulations and analyses are conducted exclusively under matched analytical dynamics, which limits the direct support for hardware robustness claims. We will revise the manuscript to clarify the scope of the results, qualify the robustness and stability statements, and discuss the implications of model mismatch for physical implementation.

read point-by-point responses

  1. Referee: [Robustness evaluation / ANOVA results] The robustness section (ANOVA on parameter variations) evaluates performance only when the test dynamics remain identical to the analytical model used for NN training; no Monte-Carlo trials or sensitivity analysis under unmodeled effects (payload variation, joint friction, or inertial mismatch) are presented, leaving the claim that the PD compensator renders the scheme “robust” for physical hardware unsupported.

    Authors: We acknowledge that the ANOVA evaluations varied parameters within the same analytical model used for NN training and testing, without introducing unmodeled effects such as friction, payload, or inertial mismatch. We will revise the robustness section to explicitly limit the claims to parametric variations under matched dynamics and to note that additional sensitivity analysis under mismatch would be required to support hardware robustness assertions. revision: yes

  2. Referee: [Comparative simulation results] All reported trajectory-tracking and torque-comparison results (including the computational-burden reduction) are generated under the exact same analytical dynamics for both training and testing; consequently the simulations do not probe the model-mismatch regime that the weakest assumption identifies as critical for the physical-robot claim.

    Authors: The referee correctly identifies that all comparative results, including computational cost reductions, were obtained under identical matched dynamics. We will revise the simulation results and discussion sections to emphasize that these outcomes hold under perfect model match and to qualify the physical-robot implications accordingly. revision: yes

  3. Referee: [Stability analysis] The analytical stability proof assumes bounded NN prediction error that the PD term can dominate, yet no explicit bound or Lipschitz analysis on the residual torque error is supplied when the inertia or Coriolis terms deviate from the training distribution; this gap directly affects the load-bearing stability guarantee.

    Authors: The stability proof relies on the assumption of bounded NN residual error under the matched model. No explicit bound or Lipschitz analysis is provided for deviations in inertia or Coriolis terms. We will revise the stability section to state the assumptions more clearly and to acknowledge that the guarantee does not extend to significant model mismatch without further analysis. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation is self-contained

full rationale

The paper generates training data from an independent analytical dynamic model of the exoskeleton, trains a parallel DNN on that data, deploys the trained network for torque prediction in simulation, and augments it with a separate PD compensator. Stability is established analytically (presumably via Lyapunov analysis on the hybrid law), and comparative simulations are run under the identical model dynamics. No claimed result reduces by construction to a fitted parameter renamed as prediction, no self-citation chain is load-bearing, and no ansatz or uniqueness theorem is smuggled in. The central claims rest on the separation between the model used for data generation and the controller implementation, which is externally verifiable and does not loop back on itself.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The central claim rests on the validity of the dynamic model for training and the assumption that simulation results translate to real-world performance.

free parameters (1)
  • Neural network architecture parameters = 4 layers, 49 neurons
    The specific number of layers and neurons was chosen for the 7 DOF robot, likely through trial or design choice.
axioms (2)
  • domain assumption The analytical dynamic model of the exoskeleton accurately captures the system behavior for generating training data
    Invoked when generating physics-based data for NN training.
  • standard math The hybrid controller combining NN prediction and PD compensation maintains stability as analytically established
    Stability is claimed to be analytically established.

pith-pipeline@v0.9.0 · 5709 in / 1372 out tokens · 34972 ms · 2026-05-24T10:42:41.251284+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

75 extracted references · 75 canonical work pages

  1. [1]

    Introduction Currently, there has been significant research focused on the advancement of exoskeleton robots. These remarkable machines have gained research attention due to their immense potential in augmenting human capabilities, revolutionize rehabilitation methods, and significantly enhance the quality of life for individuals facing mobility disabilit...

    work page

  2. [2]

    By substituting the value of the DH parameters in the general form of the homogeneous transformation matrix shown in Eqn

    Dynamic Modeling of the Exoskeleton Robot The modified DH parameter approach (Figure 2) was used to allocate link frames to individual degrees of freedom and the resultant DH parameter table (Table 1) was constructed. By substituting the value of the DH parameters in the general form of the homogeneous transformation matrix shown in Eqn. (1), the individu...

    work page

  3. [3]

    Flexion/Extension θ2 − π 2 0 0 − π 2

    work page

  4. [4]

    Knee Flexion/Extension θ4 0 0 π 2

    Internal/External rotation θ3 −l1 0 − π 2 4. Knee Flexion/Extension θ4 0 0 π 2

    work page

  5. [5]

    Ankle Dorsiflexion/Plantarflexion θ6 − π 2 0 0 π 2

    Internal rotation θ5 −l2 0 − π 2 8 6. Ankle Dorsiflexion/Plantarflexion θ6 − π 2 0 0 π 2

    work page

  6. [6]

    At any given moment the potential energy of a rigid body depends on its position Eqn

    Pronation/Supination θ7 0 a1 − π 2 I i−1T = [ cosθi −sinθi 0 αi−1 sinθicosαi−1 cosθi cosαi−1 −sinαi−1 −sinαi−1 di sinθi sinαi−1 cosθi sinαi−1 cosαi−1 cosαi−1di 0 0 0 1 ] (1) 0T1 = [ cos (θ1) −sin(θ1) 0 0 sin(θ1) cos(θ1) 0 0 0 0 1 0 0 0 0 1 ] (2) 1T2 = [ sin(θ2) cos(θ2) 0 0 0 0 0 0 cos(θ2) −sin(θ2) 1 0 0 0 0 1 ] (3) 2T3 = [ cos(θ3) −sin(θ3) 0 0 0 0 0 −l1 −...

    work page

  7. [7]

    Computed Torque control The computed torque control scheme calculates robot joints' trajectory tracking torque requirements by using robot inverse dynamics. The computed torque control system accomplishes two goals: linearize the nonlinear robot dynamics by providing the required torque to mitigate the effects of gravity and Coriolis- centrifugal force, r...

    work page

  8. [8]

    A subdivision of neural networks with more than three layers is known as a Deep neural network (DNN) [39]

    Dynamic modeling using Deep Neural Network A neural network is a collection of interconnected neurons. A subdivision of neural networks with more than three layers is known as a Deep neural network (DNN) [39]. They are increasingly used as a modeling framework for neural computations in the primate brain [40]. The main advantage of DNN is that they can le...

    work page

  9. [9]

    In this research robot’s input trajectories, user weight, and height were considered as the input of the developed deep neural network

    Input Layer: The input layer of the neural network receives the input, which is typically a vector representing the features of the data point. In this research robot’s input trajectories, user weight, and height were considered as the input of the developed deep neural network. A total of 23 inputs were considered

    work page

  10. [10]

    Each hidden layer consists of multiple neurons or nodes

    Hidden Layers: Among the input and output layers one or more hidden layers exist. Each hidden layer consists of multiple neurons or nodes. Each node in a layer is connected to all nodes in the previous layer and all nodes in the subsequent layer. There are 3 hidden layers were used for the developed deep neural network. The first hidden layer consists of ...

    work page

  11. [11]

    By adding the products of the inputs and their corresponding weights and bias, the weighted total of these inputs is obtained

    Weighted Sum: The inputs from the previous layers are multiplied by the corresponding weights in each neuron. By adding the products of the inputs and their corresponding weights and bias, the weighted total of these inputs is obtained

    work page

  12. [12]

    Typical activation functions are sigmoid, ReLU, tanh, and softmax

    Activation Function: After computing the weighted sum, an activation function is used to introduce non-linearity into the model. Typical activation functions are sigmoid, ReLU, tanh, and softmax. Based on its input, the activation function calculates the neuron's output value. In our developed deep neural network Tan hyperbolic activation function was use...

    work page

  13. [13]

    Based on the requirements, the output layer may contain one or multiple neurons

    Output Layer: The output layer receives the output of the activations from the last hidden layer. Based on the requirements, the output layer may contain one or multiple neurons. The developed deep neural network has a total of 7 neurons at the output layer. Each neuron in the output layer represents a specific class or predicts a numerical value. In our ...

    work page

  14. [14]

    In a regression problem, the output layer delivers continuous numerical value

    Output Calculation: The output layer calculates the final output based on the activations it receives. In a regression problem, the output layer delivers continuous numerical value. After completing the forward propagation backward propagation process starts in supervised learning:

    work page

  15. [15]

    The choice of loss function depends on the application, for instance, mean squared error (MSE) for regression problems or cross-entropy loss for classification problems

    Loss Calculation: Backpropagation starts with the calculation of the loss or error between the network's anticipated output and its actual output for a certain training example. The choice of loss function depends on the application, for instance, mean squared error (MSE) for regression problems or cross-entropy loss for classification problems. As the de...

    work page

  16. [16]

    The gradient of the loss function with respect to the activations of the output layer neurons is calculated to obtain this

    Output Layer Gradient: After determining the loss, the next step is to assess the relationship between the loss and the activations in the output layer. The gradient of the loss function with respect to the activations of the output layer neurons is calculated to obtain this. The specific form of the gradient calculation depends on the chosen loss function

    work page

  17. [17]

    The goal is to calculate the gradients of the loss function with respect to the weights and biases of each layer in the network

    Backpropagation through Layers: Once the gradient at the output layer is determined, it is propagated backward through the network. The goal is to calculate the gradients of the loss function with respect to the weights and biases of each layer in the network

    work page

  18. [18]

    The gradient at a particular layer depends on the gradients in the subsequent layer, weighted by the corresponding weights connecting the layers

    Gradient Calculation: In every layer, the gradients of the loss function with respect to the weights and biases are calculated using the chain rule of calculus. The gradient at a particular layer depends on the gradients in the subsequent layer, weighted by the corresponding weights connecting the layers

    work page

  19. [19]

    This is typically done using an optimization algorithm, such as gradient descent or one of its variants

    Weight Update: After computing the gradients of the loss function with respect to the weights and biases, the weights are updated to minimize the loss. This is typically done using an optimization algorithm, such as gradient descent or one of its variants. The update rule involves subtracting a fraction of the gradient from the current weight, multiplied ...

    work page

  20. [20]

    Iterative Process: Both forward and backward propagations are repeated for each training instance in the dataset. This iterative process of forward propagation, loss calculation, and backward propagation allows the model to adjust its weights and biases, gradually improving 17 its ability to make accurate predictions on unseen data. Figure 8 depicts the M...

    work page

  21. [21]

    Reference trajectories, user’s weight, and height were used as the input of the trained deep neural network

    Control algorithm based on the developed deep neural network Figure 15 shows the architecture of the hybrid controller developed to maneuver the robot according to the input reference trajectories. Reference trajectories, user’s weight, and height were used as the input of the trained deep neural network. Seven joint torques were predicted by the deep neu...

    work page

  22. [22]

    Figure 17 to Figure 22 presents the trajectory tracking performances of the developed deep neural network-based hybrid controller in simultaneous joint movements

    Simulation results analysis On this section we will simulate the human lower extremity dynamics using the developed deep neural network based hybrid controller and asses it’s performances. Figure 17 to Figure 22 presents the trajectory tracking performances of the developed deep neural network-based hybrid controller in simultaneous joint movements. Figur...

    work page

  23. [23]

    User's mass and length were varied and the effects on the trajectory tracking errors were observed to evaluate the robustness of the designed controller

    Controller robustness analysis against parameter variations ANOVA ANOVA, introduced by Ronald Fisher in 1918 [49] is a statistical technique to assess the impact of variations in system parameters on the outputs. User's mass and length were varied and the effects on the trajectory tracking errors were observed to evaluate the robustness of the designed co...

    work page 1918

  24. [24]

    Performance comparison of the developed deep neural network based controller with Computed Torque Controller, Model Reference Computed Torque Controller , Sliding Mode Controller, Adaptive controller and Linear Quadratic Regulator Control systems play a crucial role in robotic systems by providing tracking accuracy and robustness. This section would provi...

    work page

  25. [25]

    The main reason for the popularity of the deep neural network is that the trained network is computationally light in weight and can model complex dynamic systems accurately

    Conclusion Deep neural networks are increasingly being used in the field of dynamic modeling and control. The main reason for the popularity of the deep neural network is that the trained network is computationally light in weight and can model complex dynamic systems accurately. In this article, a human lower extremity exoskeleton robot dynamic model was...

    work page

  26. [26]

    Bai, and T

    Gull, M.A., S. Bai, and T. Bak, A Review on Design of Upper Limb Exoskeletons. Robotics,

    work page

  27. [27]

    Robotics and Autonomous Systems, 2015

    Yan, T., et al., Review of assistive strategies in powered lower-limb orthoses and exoskeletons. Robotics and Autonomous Systems, 2015. 64: p. 120–136

    work page 2015

  28. [28]

    Lobo-Prat, and J.M

    Rodríguez-Fernández, A., J. Lobo-Prat, and J.M. Font-Llagunes, Systematic review on wearable lower-limb exoskeletons for gait training in neuromuscular impairments. J Neuroeng Rehabil,

    work page

  29. [29]

    Hasan, S. and A. Dhingra, State of the Art Technologies for Exoskeleton Human Lower Extremity Rehabilitation Robots. Journal of Mechatronics and Robotics, 2020. 4(1)

    work page 2020

  30. [30]

    Nature Biomedical Engineering, 2023

    Siviy, C., et al., Opportunities and challenges in the development of exoskeletons for locomotor assistance. Nature Biomedical Engineering, 2023. 7(4): p. 456-472

    work page 2023

  31. [31]

    Yang, and Q

    Zhou, J., S. Yang, and Q. Xue, Lower limb rehabilitation exoskeleton robot: A review. Advances in Mechanical Engineering, 2021. 13(4): p. 16878140211011862

    work page 2021

  32. [32]

    2005: Pearson/Prentice Hall, Upper Saddle River, N.J

    Craig, J.J., Introduction to robotics : mechanics and control. 2005: Pearson/Prentice Hall, Upper Saddle River, N.J

    work page 2005

  33. [33]

    Hasan, S.K. and A.K. Dhingra, Performance verification of different control schemes in human lower extremity rehabilitation robot. Results in Control and Optimization, 2021. 4: p. 100028

    work page 2021

  34. [34]

    Hasan, S.K. and A. Dhingra, Developing a Linear Quadratic Regulator for Human Lower Extremity Exoskeleton Robot. Journal of Mechatronics and Robotics, 2022. 6(1)

    work page 2022

  35. [35]

    Hasan, S.K. and A.K. Dhingra, Development of a model reference computed torque controller for a human lower extremity exoskeleton robot. Proceedings of the Institution of Mechanical Engineers, Part I: Journal of Systems and Control Engineering, 2021. 235(9): p. 1615-1637

    work page 2021

  36. [36]

    Hasan, S.K. and A.K. Dhingra, Development of a sliding mode controller with chattering suppressor for human lower extremity exoskeleton robot. Results in Control and Optimization,

    work page

  37. [37]

    Hasan, S.K. and A.K. Dhingra, An adaptive controller for human lower extremity exoskeleton robot. Microsystem Technologies, 2021. 27(7): p. 2829-2846

    work page 2021

  38. [38]

    Hasan, S. and A.K. Dhingra, Development of a model reference computed torque controller for a human lower extremity exoskeleton robot. Proceedings of the Institution of Mechanical Engineers, Part I: Journal of Systems and Control Engineering, 2021. 235(9): p. 1615-1637

    work page 2021

  39. [39]

    Hasan, S.K. and A.K. Dhingra, 8 Degrees of freedom human lower extremity kinematic and dynamic model development and control for exoskeleton robot based physical therapy. International Journal of Dynamics and Control, 2020. 45

    work page 2020

  40. [40]

    Hasan, S.K. and A.K. Dhingra, Biomechanical design and control of an eight DOF human lower extremity rehabilitation exoskeleton robot. Results in Control and Optimization, 2022. 7: p. 100107

    work page 2022

  41. [41]

    Goodfellow, I.B.Y.C.A., Deep learning. 2016

    work page 2016

  42. [42]

    Archives of Computational Methods in Engineering, 2020

    Dargan, S., et al., A Survey of Deep Learning and Its Applications: A New Paradigm to Machine Learning. Archives of Computational Methods in Engineering, 2020. 27(4): p. 1071-1092

    work page 2020

  43. [43]

    Chen, S. and J.T. Wen, Industrial Robot Trajectory Tracking Control Using Multi-Layer Neural Networks Trained by Iterative Learning Control. Robotics, 2021. 10(1): p. 50

    work page 2021

  44. [44]

    Zhihong, S. and K. Khorasani, A neural-network-based controller for a single-link flexible manipulator using the inverse dynamics approach. IEEE Transactions on Industrial Electronics,

    work page

  45. [45]

    IEEE Transactions on Neural Networks and Learning Systems, 2021: p

    Meng, Q., et al., Motion Planning and Adaptive Neural Tracking Control of an Uncertain Two- Link Rigid-Flexible Manipulator With Vibration Amplitude Constraint. IEEE Transactions on Neural Networks and Learning Systems, 2021: p. 1-15

    work page 2021

  46. [46]

    Khorasani, and R.V

    Talebi, H.A., K. Khorasani, and R.V. Patel, Neural network based control schemes for flexible- link manipulators: simulations and experiments. Neural Netw, 1998. 11(7-8): p. 1357-1377

    work page 1998

  47. [47]

    Vandegrift, and F.L

    Yeşildirek, A., M.W. Vandegrift, and F.L. Lewis, A neural network controller for flexible-link robots. Journal of Intelligent and Robotic Systems, 1996. 17(4): p. 327-349

    work page 1996

  48. [48]

    IEEE Transactions on Industrial Informatics, 2019

    Gao, H., et al., Neural Network Control of a Two-Link Flexible Robotic Manipulator Using Assumed Mode Method. IEEE Transactions on Industrial Informatics, 2019. 15(2): p. 755-765

    work page 2019

  49. [49]

    Mechanical Systems and Signal Processing, 2020

    Yang, S., et al., An optimal fuzzy-theoretic setting of adaptive robust control design for a lower limb exoskeleton robot system. Mechanical Systems and Signal Processing, 2020. 141: p. 106706

    work page 2020

  50. [50]

    IEEE Transactions on Cybernetics, 2021

    Ren, Y., et al., Adaptive Neural-Network Boundary Control for a Flexible Manipulator With Input Constraints and Model Uncertainties. IEEE Transactions on Cybernetics, 2021. 51(10): p. 4796-4807

    work page 2021

  51. [51]

    IEEE Robotics and Automation Letters, 2021

    Lin, H., et al., Learning Deep Nets for Gravitational Dynamics With Unknown Disturbance Through Physical Knowledge Distillation: Initial Feasibility Study. IEEE Robotics and Automation Letters, 2021. 6(2): p. 2658-2665

    work page 2021

  52. [52]

    Model Learning for Robotic Manipulators using Recurrent Neural Networks

    Mukhopadhyay, R., et al. Model Learning for Robotic Manipulators using Recurrent Neural Networks. in TENCON 2019 - 2019 IEEE Region 10 Conference (TENCON). 2019

    work page 2019

  53. [53]

    IEEE Access, 2019

    Liu, N., et al., Modeling and Simulation of Robot Inverse Dynamics Using LSTM-Based Deep Learning Algorithm for Smart Cities and Factories. IEEE Access, 2019. 7: p. 173989-173998

    work page 2019

  54. [54]

    Robot Arm Dynamics Control Based on Deep Learning and Physical Simulation

    Liang, B., et al. Robot Arm Dynamics Control Based on Deep Learning and Physical Simulation. in 2018 37th Chinese Control Conference (CCC). 2018. 46

    work page 2018

  55. [55]

    Panwar, R. and N. Sukavanam, Trajectory tracking using artificial neural network for stable human-like gait with upper body motion. Neural Computing and Applications, 2020. 32(7): p. 2601-2619

    work page 2020

  56. [56]

    Narendra, K.S. and K. Parthasarathy, Identification and control of dynamical systems using neural networks. IEEE Transactions on Neural Networks, 1990. 1(1): p. 4-27

    work page 1990

  57. [57]

    Nalpantidis, and V

    Polydoros, A.S., L. Nalpantidis, and V. Krüger. Real-time deep learning of robotic manipulator inverse dynamics. in 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). 2015

    work page 2015

  58. [58]

    Zhao, and G

    Liu, C., Z. Zhao, and G. Wen, Adaptive neural network control with optimal number of hidden nodes for trajectory tracking of robot manipulators. Neurocomputing, 2019. 350: p. 136-145

    work page 2019

  59. [59]

    Dong, and C

    He, W., Y. Dong, and C. Sun, Adaptive Neural Impedance Control of a Robotic Manipulator With Input Saturation. IEEE Transactions on Systems, Man, and Cybernetics: Systems, 2016. 46(3): p. 334-344

    work page 2016

  60. [60]

    and J.J.E

    Sanner, R.M. and J.J.E. Slotine, Gaussian networks for direct adaptive control. IEEE Transactions on Neural Networks, 1992. 3(6): p. 837-863

    work page 1992

  61. [61]

    Neurocomputing, 2019

    Wu, Y., et al., Adaptive neural network control of uncertain robotic manipulators with external disturbance and time-varying output constraints. Neurocomputing, 2019. 323: p. 108-116

    work page 2019

  62. [62]

    Liu, and G

    Zhu, Q., Y. Liu, and G. Wen, Adaptive neural network control for time-varying state constrained nonlinear stochastic systems with input saturation. Information Sciences, 2020. 527: p. 191-209

    work page 2020

  63. [63]

    Friction in contact between rotating bodies

    Mathworks.com. Friction in contact between rotating bodies. 2020 07/01/2020]; Available from: https://www.mathworks.com/help/physmod/simscape/ref/rotationalfriction.html

    work page 2020

  64. [64]

    Nature Communications, 2020

    Mehrer, J., et al., Individual differences among deep neural network models. Nature Communications, 2020. 11(1): p. 5725

    work page 2020

  65. [65]

    Deep Learning Vs Neural Networks – What’s The Difference ? ; Available from: https://bernardmarr.com/deep-learning-vs-neural-networks-whats-the-difference/

    Marr, B. Deep Learning Vs Neural Networks – What’s The Difference ? ; Available from: https://bernardmarr.com/deep-learning-vs-neural-networks-whats-the-difference/

    work page

  66. [66]

    IBM. AI vs. Machine Learning vs. Deep Learning vs. Neural Networks Available from: https://www.ibm.com/cloud/blog/ai-vs-machine-learning-vs-deep-learning-vs-neural-networks

    work page

  67. [67]

    2023 [cited 2023 06/04/2023]; Available from: https://www.tutorialspoint.com/python_deep_learning/python_deep_learning_deep_neural_netwo rks.htm

    Deep Neural Networks. 2023 [cited 2023 06/04/2023]; Available from: https://www.tutorialspoint.com/python_deep_learning/python_deep_learning_deep_neural_netwo rks.htm

    work page 2023

  68. [68]

    06/012023 [cited 2023; Available from: https://www.sas.com/en_us/insights/analytics/neural-networks.html

    Artificial Neural Networks: What they are & why they matter. 06/012023 [cited 2023; Available from: https://www.sas.com/en_us/insights/analytics/neural-networks.html

    work page 2023

  69. [69]

    E, B. and D. B. Performance Analysis and CPU vs GPU Comparison for Deep Learning. in 2018 6th International Conference on Control Engineering & Information Technology (CEIT). 2018

    work page 2018

  70. [70]

    Bengio, Y., Learning Deep Architectures for AI. Found. Trends Mach. Learn., 2009. 2(1): p. 1– 127. 47

    work page 2009

  71. [71]

    [cited 2023 06/01/2023]; Available from: https://www.mathworks.com/help/deeplearning/ref/feedforwardnet.html

    Generate feedforward neural network. [cited 2023 06/01/2023]; Available from: https://www.mathworks.com/help/deeplearning/ref/feedforwardnet.html

    work page 2023

  72. [72]

    The Levenberg-Marquardt algorithm for nonlinear least squares curve-fitting problems

    Gavin, H.P. The Levenberg-Marquardt algorithm for nonlinear least squares curve-fitting problems. 2019

    work page 2019

  73. [73]

    Levenberg-Marquardt backpropagation

    Inc., T.M. Levenberg-Marquardt backpropagation. 2023 [cited 2023 07/06/2023]; Available from: https://www.mathworks.com/help/deeplearning/ref/trainlm.html

    work page 2023

  74. [74]

    Kotz and N.L

    Fisher, R.A., Statistical Methods for Research Workers, in Breakthroughs in Statistics: Methodology and Distribution, S. Kotz and N.L. Johnson, Editors. 1992, Springer New York: New York, NY. p. 66-70

    work page 1992

  75. [75]

    Hasan, S.K. and A.K. Dhingra, Developing a Linear Quadratic Regulator for Human Lower Extremity Exoskeleton Robot. Journal of Mechatronics and Robotics, 2022. 48 List of Figures Figure 1 Mechatronics architecture of a robot .............................................................................................. 3 Figure 2 Link Frame assignment based...

    work page 2022