Pith. sign in

REVIEW 4 major objections 5 minor 28 references

Deep Reinforcement Learning-Based Motion Planning and PDE Control for Flexible Manipulators

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read DRL-based motion planning plus a nonlinear PDE controller can track flexible-arm angles while suppressing endpoint vibration.

desk verdict Real hardware and a plausible planner-controller integration, but the central Lyapunov stability theorem is wrong as stated and the evaluation has a self-training bias. read the letter →

arxiv 2506.08639 v1 pith:5NLKIY2Y submitted 2025-06-10 cs.RO cs.SYeess.SY

Amir Hossein Barjini , Seyed Adel Alizadeh Kolagar , Sadeq Yaqubi , Jouni Mattila This is my paper
classification cs.ROcs.SYeess.SY
keywords flexiblemanipulatorsdeepreinforcementlearningsoftactor-criticPDEcontrolLyapunovstabilityvibrationsuppressionmotionplanningunderactuatedsystems
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

This paper argues that for a flexible robot arm, the reference trajectory itself is a vibration-suppression lever: a bad path excites bending modes that a single base torque cannot cancel. To use that lever, the authors pair a soft actor-critic (SAC) reinforcement-learning planner, which outputs desired joint velocities, with a nonlinear PDE-based controller that converts those velocities into one torque at the base. The claimed result is precise angle tracking together with far smaller payload vibration than PID or PDE-only control of a cubic-polynomial path. Simulations and experiments on a 4.5 m hydraulically actuated flexible link support the claim, with endpoint elastic deviation kept near 3–6 mm. The paper further asserts exponential closed-loop stability through Lyapunov analysis; whether that proof step holds is the load-bearing premise recorded below.

What carries the argument

The load-bearing mechanism is a high-level/low-level loop. The high level is a soft actor-critic policy $\pi_\phi(a_t|s_t)$ over the state $s_t = \langle e_T(t), \theta(t), \dot{\theta}(t), \tau(t), \omega(L,t), \dot{\omega}(L,t)\rangle$, with actions chosen as normalized joint velocities; the low level is the nonlinear PDE controller of equation (17), which feeds beam boundary information $\omega''(0,t)$ and $\omega(L,t)$ into the torque. To keep the PDE model tractable, the paper applies the homogenizing transformation $z(x,t) = \omega(x,t) + \nu(x,t)$ of equation (9), with $\nu$ given in closed form so that the payload boundary condition becomes homogeneous and an assumed-mode solution can be used. The intended exponential-stability argument is carried by the cross term $\alpha I_m e\dot{e}$ in the Lyapunov function: the paper uses it to convert a stability margin into the decay inequality $\dot{V}<-\lambda V$.

What would settle it

Simulate a raising-lowering maneuver with the paper's gains and compute $\dot{V}$ from equation (24) pointwise; if $\dot{V}(t)>0$ at any time when $e(t)$ and $\dot{e}(t)$ have opposite signs, the asserted inequality $\dot{V}<-\lambda V$ fails and Theorem 1's proof does not establish exponential stability.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that trajectory planning, not feedback alone, is the deciding factor for vibration suppression in an underactuated flexible arm. With a single input torque, the nonlinear PDE controller $\tau(t) = K_p e(t) + K_d \dot{e}(t) - H(t)$, with $H(t) = EI\,\omega''(0,t) - \tfrac{1}{2}mgL\cos\theta(t) + Mg\,\omega(L,t)\sin\theta(t) - I_m\ddot{\theta}_d(t)$, can track the desired angle but leaves the bending mode largely untouched. The proposed fix is a two-level loop: a SAC policy trained with domain randomization emits normalized joint velocities $\dot{\theta}_d = \dot{\theta}_{\max} a_t$, integration produces $\theta_d(t)$, and the PDE controller realizes the tracking torque. In simulation the combined scheme lowers the tip-velocity RMSE to $1.2906\times 10^{-4}$ m/s, roughly an order of magnitude below the PDE-with-cubic-trajectory baseline, and in hardware it keeps endpoint deflection within about 3–6 mm. Theorem 1 states that the same control law guarantees exponential tracking; the proof rests on the Lyapunov function $V = \tfrac{1}{2}K_p e^2 + \tfrac{1}{2}I_m\dot{e}^2 + \alpha I_m e\dot{e}$ and the asserted decay $\dot{V} < -\lambda V$.

Load-bearing premise

The stability proof assumes the mixed error term in the Lyapunov function can never work against decay; because that term changes sign, the claimed negative bound is not guaranteed by the stated conditions.

Editorial extensions

If this is right

  • For underactuated flexible arms, shaping the reference path becomes a first-class vibration-suppression tool rather than a separate planning concern.
  • A 10 Hz model-free planner can be safely coupled to a 1 kHz model-based controller, so the scheme is real-time compatible with hydraulic actuation.
  • Because the planner needs no endpoint actuation, the architecture transfers to arms where only a base torque is available.
  • Domain randomization over link parameters produces a policy that carries over from simulation to hardware without retraining, as reflected in the reported 3–6 mm endpoint amplitudes.

Reading between the lines

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

  • A direct extension would be a multi-link flexible arm, with per-link versions of the tip-velocity penalty $W_{\dot{\omega}}|\dot{\omega}(L,t)|$; the paper does not test that setting.
  • The ±10% domain-randomization envelope defines the planner's expected operating range; retraining would likely be needed for substantially different payloads or link stiffness.
  • The reward weights are manually tuned; scheduling them during training could be an automatic way to balance reaching speed against vibration, but the paper fixes them.
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 / 5 minor

Summary. The manuscript proposes a two-layer framework for a single-link flexible manipulator: a soft actor-critic (SAC) motion planner generates a desired joint-angle trajectory, and a nonlinear PDE controller (Eq. 17) tracks that trajectory using a single base torque. The flexible link is modeled as an Euler-Bernoulli beam with a payload, leading to the PDE system in Eqs. (4)-(8), with a transformation in Eqs. (9)-(16) used to homogenize non-homogeneous boundary conditions. The authors claim exponential closed-loop stability via Lyapunov analysis (Theorem 1) and validate the approach in MATLAB simulations and on a 4.5 m hydraulically actuated flexible manipulator, reporting improved tracking and reduced endpoint vibration compared with PID control and cubic-polynomial-trajectory baselines.

Significance. If fully established, the proposed division of labor between learned motion planning and model-based PDE control would be a useful contribution for underactuated flexible robots, and the hardware experiments on a large-scale manipulator are a valuable practical element. The paper also provides an explicit PDE model with non-homogeneous boundary conditions and uses domain randomization during DRL training, which are positive features. However, the central theoretical claim of exponential closed-loop stability is not proven as stated, and the experimental evaluation is too thin to fully support the performance claims. The framework may be salvageable with a corrected theorem, additional evaluation metrics, and repeated experimental trials, but the current manuscript does not establish its main contributions.

major comments (4)
  1. [III.A, Eqs. (24)-(30)] The step from Eq. (28) to Eq. (30) is invalid. The cross term -λ3(αIme e_dot) in Eq. (28) is sign-indefinite because e e_dot can be negative. Choosing λ < min{λ1, λ2, λ3} in Eq. (29) only controls the coefficients of the positive quadratic terms; it does not control the cross term. For states with e e_dot < 0, the inequality V_dot < -λV can fail. Concretely, with α=0.9, Im=1, Kp=Kd=0.95, λ=0.09, e=1, and e_dot=-10, one obtains V_dot + λV ≈ 6.43 > 0, contradicting Eq. (30). The exponential-stability conclusion in Theorem 1 therefore does not follow from the given Lyapunov argument.
  2. [III.A, Theorem 1] The Lyapunov function V in Eq. (19) contains only the tracking error e and its derivative e_dot; it contains no measure of the beam state ω(x,t) or its energy. Because H(t) in Eq. (18) cancels all beam terms in the e_dot dynamics, the argument at best proves exponential convergence of e to zero. It cannot establish exponential stability of the full infinite-dimensional closed loop. In fact, if θ = θd, e = 0, e_dot = 0, but the beam has nonzero initial deflection, the control torque τ = -H leaves θ_ddot = 0, and the undamped Euler-Bernoulli equation (5) with boundary conditions (7) permits persistent oscillation. Exponential stability of the full closed-loop system is therefore false as stated unless structural damping is added to the model. The theorem and contribution 1 must be reformulated, or the model must include damping.
  3. [III.B and IV.C, Eq. (38)] The principal evaluation metric for vibration suppression, the RMSE of ω_dot(L,t), is exactly the quantity penalized by the reward term W_ωdot |ω_dot(L,t)| in Eq. (38). The DRL planner is directly optimized on this quantity, so the reported improvement over the CPT baseline is partly by construction. To substantiate the claim of superior vibration suppression, the authors should report evaluation metrics that are not part of the reward, such as peak elastic deflection, settling time, or a frequency-domain vibration metric, and compare against a non-learning trajectory-optimization baseline or an ablation without the tip-velocity penalty.
  4. [V, Fig. 7] The experimental validation consists of a single run for each condition. No repeated trials, error bars, or statistical measures are reported, so the claimed differences in vibration amplitude (e.g., 'below 3 mm when raising and 6 mm when lowering') cannot be assessed against run-to-run variability. The robustness study in Table III covers parameter uncertainty for the PDE controller only, not for the DRL planner; a comparable uncertainty or repeatability study is needed for the combined method.
minor comments (5)
  1. [Table III] The heading 'RSME' should be 'RMSE'.
  2. [Fig. 7c and Section V.C] The text reports vibration amplitudes of 3 mm and 6 mm, but the vertical axis shows absolute elastic deviations around -0.23 m; please specify the baseline used to define amplitude (e.g., static deflection) and clarify whether the quoted values are deviations from that baseline.
  3. [Section IV.C] The statement that vibrations are kept under 1.5 cm should be reconciled with the axis range in Fig. 5b; clarify whether this refers to oscillation amplitude around the static deflection rather than absolute elastic deviation.
  4. [Eq. (38)] The reward terms are described as penalties but are written with negative weights; a short explanation of the sign convention would improve readability.
  5. [General] The manuscript does not mention software or data availability; releasing the training code and recorded experimental data would improve reproducibility, which is especially important for a learning-based method.

Circularity Check

1 steps flagged · score 6.0 of 10

DRL vibration-suppression result is partly by construction: the training reward includes the exact tip-velocity metric later used as the main evaluation criterion.

  1. fitted input called prediction [Section III-B, Eq. (38); Section IV-C, RMSE comparison paragraph]
    "Rt = We|eT (t)| + W ˙θ| ˙θ (t)| + W ˙ω| ˙ω(L,t)| + Rreach + R f ailure. ... The third term, W ˙ω| ˙ω(L,t)|, addresses tip vibration by penalizing excessive tip velocity, helping to reduce oscillations inherent in flexible manipulators. ... To evaluate vibration suppression at the flexible link tip, we compared the RMSE of ˙ω(L,t). Our PDE controller with DRL motion planner achieved the lowest RMSE of 1.2906×10−4 m/s, compared to 1.6×10−3 m/s for PDE controller with CPT method and 2.0807×10−4 m/s for PID controller with CPT method, confirming its superior vibration mitigation."

    The SAC agent is trained to maximize a reward that includes the penalty W_ωdot|ωdot(L,t)|, i.e., the magnitude of the tip velocity at every step. The paper's principal vibration-suppression evidence is then the RMSE of ωdot(L,t) over the trajectory. Because the optimized policy is explicitly rewarded for minimizing the very signal whose RMSE is reported, the comparison is a restatement of the training objective rather than an independent test. Baselines trained without this objective are expected to fare worse on this metric, so the claimed vibration reduction is partly by construction.

full rationale

The PDE controller derivation itself is self-contained: substituting (17) into (23) cancels the beam terms through H, and the remaining calculation is an algebraic Lyapunov argument over e and edot. The homogenization transformation (9)-(16) is cited to [14], which overlaps with the authors, but the equations are written out and checkable, so that self-citation is not load-bearing. The mathematical flaw in the λ step of Theorem 1 involving the sign-indefinite αIm e edot term is a correctness defect, not a circularity, so it is not scored here. The one genuine circular element is the DRL planner: Eq. (38) includes W_ωdot|ωdot(L,t)| as a reward term, and Section IV-C evaluates success by the RMSE of the same ωdot(L,t) signal. That makes the headline claim that the DRL planner suppresses tip vibrations partly by construction, although the real-world experiments and tracking results give the framework some independent content. Hence partial circularity rather than full circularity.

Assumptions & free parameters 9 free parameters · 5 assumptions · 0 invented entities

The central claims rest on manual parameter choices (controller gains and reward weights), a truncated beam model, and an unproven Lyapunov inequality. No new physical entities are introduced. The largest burden is the stability proof gap, followed by the reward-metric coupling between the DRL objective and the evaluation.

free parameters (9)
  • Kp (PDE controller proportional gain) = 12000 (simulation; experimental value not stated)
    Manually selected gain for the tracking controller; not derived from the model or an optimization. Used in equation (17).
  • Kd (PDE controller derivative gain) = 15000 (simulation; experimental value not stated)
    Manually selected derivative gain in equation (17).
  • Reward weight We = -5e-3
    Manually tuned weight for the angle-error term in the reward function (38).
  • Reward weight W_theta_dot = -1e-3
    Manually tuned penalty on joint velocity in the reward function (38).
  • Reward weight W_omega_dot = -3e-1
    Manually tuned penalty on tip velocity in the reward function (38); this is also the main evaluation metric.
  • Rreach = +200
    Manually chosen positive reward for reaching the target under tolerance (40).
  • Rfailure = -200
    Manually chosen penalty for violating angle limits (41).
  • Maximum joint velocity theta_dot_max = 5 deg/s
    Manually specified action scale (36), chosen based on physical limitations; affects the planner's reachable trajectories.
  • Domain randomization range = ±10%
    Perturbation range applied to [L, m, M, Im, rho A, EI] during training; a manual choice that influences sim-to-real transfer.
assumptions (5)
  • domain assumption Euler-Bernoulli beam theory with small deformations and linear elastic material holds for the 4.5 m steel link
    The PDE model (4)-(7) is built on Euler-Bernoulli assumptions; the real link has finite thickness and possible shear deformation, which are neglected.
  • domain assumption The Assumed Mode Method with three modes accurately represents the infinite-dimensional beam dynamics in simulation
    Section IV states 'considering three modes' without convergence analysis; the controller design uses the full PDE while the simulation uses a truncated model.
  • ad hoc to paper Equation (30), Vdot < -lambda V, holds for all e and edot with lambda < min(lambda1, lambda2, lambda3)
    This step in the Lyapunov proof is not justified because the cross term alpha Im e edot is sign-indefinite; the comparison fails when the cross term is negative.
  • standard math The homogenization transformation z = omega + nu from reference [14] is correct and applicable to the non-homogeneous boundary condition (8)
    The transformation (9)-(16) is cited from the authors' prior work [14] and used without independent derivation in this paper.
  • domain assumption The simulation environment with domain randomization transfers to the real hydraulic manipulator without fine-tuning
    The DRL planner is trained in MATLAB Simulink and applied directly to the real robot; the paper does not verify the fidelity of the simulation to the real actuator dynamics or sensor noise.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Deep Reinforcement Learning-Based Motion Planning and PDE Control for Flexible Manipulators." pith.science (2026). https://pith.science/paper/5NLKIY2Y

@misc{pith2026250608639,
  author       = {Pith},
  title        = {Pith review of: Deep Reinforcement Learning-Based Motion Planning and PDE Control for Flexible Manipulators},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5NLKIY2Y}},
  note         = {Machine review of arXiv:2506.08639}
}
read the original abstract

This article presents a motion planning and control framework for flexible robotic manipulators, integrating deep reinforcement learning (DRL) with a nonlinear partial differential equation (PDE) controller. Unlike conventional approaches that focus solely on control, we demonstrate that the desired trajectory significantly influences endpoint vibrations. To address this, a DRL motion planner, trained using the soft actor-critic (SAC) algorithm, generates optimized trajectories that inherently minimize vibrations. The PDE nonlinear controller then computes the required torques to track the planned trajectory while ensuring closed-loop stability using Lyapunov analysis. The proposed methodology is validated through both simulations and real-world experiments, demonstrating superior vibration suppression and tracking accuracy compared to traditional methods. The results underscore the potential of combining learning-based motion planning with model-based control for enhancing the precision and stability of flexible robotic manipulators.

Figures

Figures reproduced from arXiv: 2506.08639 by the authors.

Figure 1
Figure 1. Schematic View of the Flexible Manipulator. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Schematic View of the Proposed High-Level Motion Pla [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Comparison of the PID Controller and the Proposed PDE [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Rewards in the Training Process. The simulation results using the proposed PDE controller and the DRL motion planner are illustrated in [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Simulation Results of the Proposed PDE Controller Us [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Experimental Platform of the Flexible Manipulator. [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Experimental Results. (a) Angular Position Using th [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

28 extracted references · 27 canonical work pages

  1. [13]

    Deep reinforcement learning with reward shapi ng for tracking control and vibration suppression of flexible link manipulator,

    J. K. Viswanadhapalli, V . K. Elumalai, S. Shivram, S. Sh ah, and D. Mahajan, “Deep reinforcement learning with reward shapi ng for tracking control and vibration suppression of flexible link manipulator,” Applied Soft Computing , vol. 152, p. 110756, 2024

    work page 2024

  2. [17]

    Reinforcement learning control of a flexible two-link manipulator: An experimental investiga- tion,

    W. He, H. Gao, C. Zhou, C. Y ang, and Z. Li, “Reinforcement learning control of a flexible two-link manipulator: An experimental investiga- tion,” IEEE Transactions on Systems, Man, and Cybernetics: System s, vol. 51, no. 12, pp. 7326–7336, 2020

    work page 2020

  3. [1]

    A critical review of control techniques for flexible and rigid link manipulators,

    E. A. Alandoli and T. S. Lee, “A critical review of control techniques for flexible and rigid link manipulators,” Robotica, vol. 38, no. 12, pp. 2239–2265, 2020

    work page 2020

  4. [2]

    A variable length, variable stiffness flexible inst rument for transoral robotic surgery,

    J. Chen, Q. Ding, W. Y an, K. Y an, J. Chen, J. Y .-K. Chan, and S. S. Cheng, “A variable length, variable stiffness flexible inst rument for transoral robotic surgery,” IEEE Robotics and Automation Letters , vol. 7, no. 2, pp. 3835–3842, 2022

    work page 2022

  5. [3]

    Designing, protot yping, and testing a flexible suturing robot for transanal endoscopic m icrosurgery,

    Y . Hu, W. Li, L. Zhang, and G.-Z. Y ang, “Designing, protot yping, and testing a flexible suturing robot for transanal endoscopic m icrosurgery,” IEEE Robotics and Automation Letters , vol. 4, no. 2, pp. 1669–1675, 2019

    work page 2019

  6. [4]

    Revie w of modelling and control of flexible-link manipulators,

    M. Sayahkarajy, Z. Mohamed, and A. A. Mohd Faudzi, “Revie w of modelling and control of flexible-link manipulators,” Proceedings of the Institution of Mechanical Engineers, Part I: Journal of Sys tems and Control Engineering, vol. 230, no. 8, pp. 861–873, 2016

    work page 2016

  7. [5]

    Model predictive control applied to different time- scale dynamics of flexible joint robots,

    M. Iskandar, C. van Ommeren, X. Wu, A. Albu-Sch¨ affer, an d A. Diet- rich, “Model predictive control applied to different time- scale dynamics of flexible joint robots,” IEEE Robotics and Automation Letters , vol. 8, no. 2, pp. 672–679, 2022

    work page 2022

  8. [6]

    Dyna mic modelling and vibration suppression of a single-link flexib le manipulator with two cables,

    L. Tang, M. Gouttefarde, H. Sun, L. Yin, and C. Zhou, “Dyna mic modelling and vibration suppression of a single-link flexib le manipulator with two cables,” Mechanism and Machine Theory , vol. 162, p. 104347, 2021

    work page 2021

Show all 28 references
  1. [7]

    Visual servoing of fl exible-link manipulators by considering vibration suppression withou t deformation measurements,

    K. Li, H. Wang, X. Liang, and Y . Miao, “Visual servoing of fl exible-link manipulators by considering vibration suppression withou t deformation measurements,” IEEE Transactions on Cybernetics , vol. 52, no. 11, pp. 12 454–12 463, 2021

    2021

  2. [8]

    Backstepping cont rol for vibration suppression of 2-d euler–bernoulli beam based on nonlinear saturation compensator,

    Z. Jing, Y . Ma, X. Wu, X. He, and Y . Sun, “Backstepping cont rol for vibration suppression of 2-d euler–bernoulli beam based on nonlinear saturation compensator,” IEEE Transactions on Systems, Man, and Cybernetics: Systems , vol. 53, no. 5, pp. 2562–2571, 2022

    2022

  3. [9]

    Vibration control of a flexible beam wit h output constraint,

    W. He and S. S. Ge, “Vibration control of a flexible beam wit h output constraint,” IEEE Transactions on Industrial Electronics , vol. 62, no. 8, pp. 5023–5030, 2015

    2015

  4. [10]

    Boundary disturbance obse rver- based control of a vibrating single-link flexible manipulat or,

    Z. Zhao, X. He, and C. K. Ahn, “Boundary disturbance obse rver- based control of a vibrating single-link flexible manipulat or,” IEEE Transactions on Systems, Man, and Cybernetics: Systems , vol. 51, no. 4, pp. 2382–2390, 2019

    2019

  5. [11]

    Boundary control for a flexib le manipulator based on infinite dimensional disturbance observer,

    T. Jiang, J. Liu, and W. He, “Boundary control for a flexib le manipulator based on infinite dimensional disturbance observer,” Journal of sound and Vibration, vol. 348, pp. 1–14, 2015

    2015

  6. [12]

    Deep learning-based deflection correction and end-point contro l of heavy-duty vertical single-link flexible manipulators,

    A. H. Barjini, S. Y aqubi, S. M. Tahamipour-Z, and J. Matt ila, “Deep learning-based deflection correction and end-point contro l of heavy-duty vertical single-link flexible manipulators,” in 2024 IEEE International Conference on Robotics and Biomimetics (ROBIO) . IEEE, 2024, pp...

    2024

  7. [14]

    Semi-an alytical de- sign of pde endpoint controller for flexible manipulator wit h non- homogenous boundary conditions,

    S. Y aqubi, S. M. Tahamipour-Z, and J. Mattila, “Semi-an alytical de- sign of pde endpoint controller for flexible manipulator wit h non- homogenous boundary conditions,” IEEE Transactions on Automation Science and Engineering , 2023

    2023

  8. [15]

    Boundary force control of t wo one-link flexible manipulator to accomplish safe grasping task,

    H. Sayyaadi and M. Hejrati, “Boundary force control of t wo one-link flexible manipulator to accomplish safe grasping task,” in 29th Annu. Int. Conf. Iran. Assoc. Mech. Eng. 8th Int. Conf. Therm. Powe r Plants Ind., 2021

    2021

  9. [16]

    Neural network contro l of a two- link flexible robotic manipulator using assumed mode method ,

    H. Gao, W. He, C. Zhou, and C. Sun, “Neural network contro l of a two- link flexible robotic manipulator using assumed mode method ,” IEEE Transactions on Industrial Informatics , vol. 15, no. 2, pp. 755–765, 2018

    2018

  10. [18]

    Design of a sliding mode controller for suppressing coupled axial & torsional v ibrations in horizontal drill strings using extended kalman filter,

    A. H. Barjini, M. Khoshnazar, and H. Moradi, “Design of a sliding mode controller for suppressing coupled axial & torsional v ibrations in horizontal drill strings using extended kalman filter,” Journal of Sound and Vibration, vol. 586, p. 118477, 2024

    2024

  11. [19]

    A review of r ecent trend in motion planning of industrial robots,

    M. G. Tamizi, M. Y aghoubi, and H. Najjaran, “A review of r ecent trend in motion planning of industrial robots,” International Journal of Intelligent Robotics and Applications , vol. 7, no. 2, pp. 253–274, 2023

    2023

  12. [20]

    A review of rob ot learning for manipulation: Challenges, representations, and algor ithms,

    O. Kroemer, S. Niekum, and G. Konidaris, “A review of rob ot learning for manipulation: Challenges, representations, and algor ithms,” Journal of machine learning research , vol. 22, no. 30, pp. 1–82, 2021

    2021

  13. [21]

    R. S. Sutton, A. G. Barto et al., Reinforcement learning: An introduction. MIT press Cambridge, 1998, vol. 1, no. 1

    1998

  14. [22]

    A review of deep reinforcement learning approaches for smart manufact uring in industry 4.0 and 5.0 framework,

    A. del Real Torres, D. S. Andreiana, ´A. Ojeda Rold´ an, A. Hern´ andez Bustos, and L. E. Acevedo Galicia, “A review of deep reinforcement learning approaches for smart manufact uring in industry 4.0 and 5.0 framework,” Applied Sciences , vol. 12, no. 23, p. 12377, 2022

    2022

  15. [23]

    Deep reinf orcement learning for robotic manipulation with asynchronous off-p olicy updates,

    S. Gu, E. Holly, T. Lillicrap, and S. Levine, “Deep reinf orcement learning for robotic manipulation with asynchronous off-p olicy updates,” in 2017 IEEE international conference on robotics and automat ion (ICRA). IEEE, 2017, pp. 3389–3396

    2017

  16. [24]

    Deep r einforcement learning-based path planning for multi-arm manipulators w ith periodi- cally moving obstacles,

    E. Prianto, J.-H. Park, J.-H. Bae, and J.-S. Kim, “Deep r einforcement learning-based path planning for multi-arm manipulators w ith periodi- cally moving obstacles,” Applied Sciences, vol. 11, no. 6, p. 2587, 2021

    2021

  17. [25]

    Soft actor-critic: Off-policy maximum entropy deep reinforcement learning with a stochastic actor,

    T. Haarnoja, Zhou et al., “Soft actor-critic: Off-policy maximum entropy deep reinforcement learning with a stochastic actor,” pp. 1 861–1870, 2018

    2018

  18. [26]

    End-to-e nd training of deep visuomotor policies,

    S. Levine, C. Finn, T. Darrell, and P . Abbeel, “End-to-e nd training of deep visuomotor policies,” Journal of Machine Learning Research , vol. 17, no. 39, pp. 1–40, 2016

    2016

  19. [27]

    Integrat ing deeprl with robust low-level control in robotic manipulators for n on-repetitive reaching tasks,

    M. H. Shahna, S. A. A. Kolagar, and J. Mattila, “Integrat ing deeprl with robust low-level control in robotic manipulators for n on-repetitive reaching tasks,” in 2024 IEEE International Conference on Mechatronics and Automation (ICMA) . IEEE, 2024, pp. 329–336

    2024

  20. [28]

    On the role of the action space in robot manipulation learning and sim-to- real transfer,

    E. Aljalbout, F. Frank, M. Karl, and P . van der Smagt, “On the role of the action space in robot manipulation learning and sim-to- real transfer,” IEEE Robotics and Automation Letters , 2024

    2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.