Pith. sign in

REVIEW 5 major objections 5 minor 39 references

Cross-Platform Learnable Fuzzy Gain-Scheduled Proportional-Integral-Derivative Controller Tuning via Physics-Constrained Meta-Learning and Reinforcement Learning Adaptation

T0 review · 5 major / 5 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read A hierarchical pipeline—meta-learned fuzzy PID initialization from robot physical features plus a lightweight RL refinement stage—cuts tracking error by up to 80.4% on high-load joints and 19.2% under parameter uncertainty.

desk verdict Plausible pipeline, but the cross-platform claim is not actually tested: both test robots are also meta-training bases, and missing baselines plus internal contradictions make the empirical case unverifiable as written. read the letter →

arxiv 2511.06500 v2 pith:5IGTQ2MF submitted 2025-11-09 cs.RO

JiaHao Wu , ShengWen Yu This is my paper
classification cs.RO
keywords fuzzyPIDgainschedulingmeta-learningreinforcementlearningcontrollerauto-tuningcross-platformgeneralizationphysics-baseddataaugmentationrobotcontrol
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 tries to establish that the long-standing pain point of PID tuning—gains that work on one robot usually do not transfer to another—can be solved by splitting the problem in two. A meta-learner first reads a 10-dimensional summary of a robot's physics (mass, degrees of freedom, inertia, link lengths, center of mass, friction) and predicts a complete set of fuzzy gain-scheduling parameters, so a new platform gets a strong controller in under a millisecond. A reinforcement-learning stage then adjusts only the per-joint input scalings of the fuzzy system during deployment, converging in about ten minutes. On two morphology-different unseen platforms, the pipeline reduces mean tracking error by up to 16.6% overall and 80.4% on a high-inertia joint, and by 19.2% under parameter uncertainty. The paper's deeper claim is a design law: RL refinement pays off when the meta-initialized controller has localized weak joints, but is nearly useless when the baseline is already uniformly strong.

What carries the argument

The load-bearing object is the learnable fuzzy gain-scheduled PID (LF-PID), a PID whose gains are scheduled online by a Takagi–Sugeno fuzzy system (each rule returns an affine function of the error, error-rate, and error-integral inputs). Two ideas carry the argument: shared membership partitions, which keep error semantics comparable across joints and platforms, and a compact RL adaptation space—only the per-joint input scaling factors are retuned, while the fragile fuzzy consequents are frozen at the meta-learned values. That split is what makes the whole pipeline sample-efficient: the meta-network gives a good initialization from 232 physics-constrained virtual robots, and the RL stage fi

What would settle it

A reader could falsify the feature-sufficiency claim by training the meta-network on only one base morphology (e.g., 7-DOF arm variants) and testing on the other morphology (12-DOF quadruped): if zero-shot prediction is no better than random and RL cannot recover within the 1M-step budget, the 10D features are not carrying the generalization. Another concrete check: on the meta-initialized baseline, compute the coefficient of variation of per-joint errors across many seeds; if CV < 0.2 yet RL still gives >5% aggregate improvement, the optimization ceiling effect fails.

Watch

Extended reading notes

Core claim

The paper's central claim is that a learnable fuzzy gain-scheduled PID (LF-PID) can be tuned across robot platforms by a hierarchical meta-learning plus RL scheme. The LF-PID uses shared fuzzy membership partitions so that 'small error' means the same thing across robots, while per-joint scaling factors and Takagi–Sugeno consequent parameters adapt the PID gains online. A neural network maps a 10D physical feature vector to all LF-PID parameters, trained on 232 physically valid virtual robots generated by bounded perturbations of mass, inertia, and friction; a policy-gradient RL stage then refines only the 3n input scalings, leaving base gains and consequents fixed. In cross-platform tests t

Load-bearing premise

The 10-dimensional physical feature vector (DOF, total mass, inertia tensor components, link lengths, COM, friction) is assumed to be a sufficient summary of a robot's dynamics, so that a network trained on perturbed arm and quadruped variants predicts near-optimal LF-PID parameters for a genuinely new morphology.

Editorial extensions

If this is right

  • If correct, a new robot's fuzzy PID parameters can be initialized in about 0.8 ms from a 10D physical feature vector, and deployment-specific refinement costs roughly 10 minutes—instead of 40–120 hours of expert tuning or 30–60 minutes of per-platform re-optimization.
  • The optimization ceiling effect gives a practical go/no-go test: measure the per-joint error distribution of the meta-initialized controller; if it is uniform with low coefficient of variation, skip RL and keep the initialization.
  • Physics-constrained augmentation is load-bearing: removing it raises meta-learning prediction error from 3.33% to 31.2%, so bounded perturbations plus validity checks are what make 232 samples enough.
  • The 19.2% gain under mass/inertia/friction uncertainty suggests the method can absorb model mismatch, which is the usual gap between simulation and deployment.
  • All results are simulation-based; the paper itself lists sensor noise, unmodeled friction, latency, and safety as open items before real-world use.

Reading between the lines

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

  • The strongest untested consequence is that the same 10D feature-to-parameter map extends beyond the arm-quadruped pair—to humanoids, mobile manipulators, or underwater vehicles—and this is probably testable by retraining only the virtual-robot synthesis, not the architecture.
  • If the ceiling effect is real, then the natural next step is an adaptive meta-learner that detects per-joint error heterogeneity during training and biases its predictions to be less uniform, which would reduce the need for the RL stage.
  • A cheap monitoring metric the authors only gesture at: compute the coefficient of variation of per-joint MAE after meta-initialization; using it as a real-time trigger for RL could save compute on easy platforms.
  • Hardware validation with encoder noise and control latency is the decisive experiment; the conservative perturbation ranges and position-control abstraction make transfer plausible, but nothing in the simulation guarantees it.
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

5 major / 5 minor

Summary. The paper proposes a hierarchical framework for tuning a learnable fuzzy gain-scheduled PID (LF-PID) controller across robot platforms. The controller maintains shared fuzzy membership partitions while learning per-joint scaling and Takagi–Sugeno consequent parameters. A physics-constrained augmentation stage generates 232 virtual robot variants from three base simulated platforms (Franka Panda, KUKA LBR iiwa, Laikago) by bounded perturbations of mass, inertia, link lengths, and friction. A meta-learning network maps a 10-dimensional robot-feature vector to the full LF-PID parameter vector, providing zero-shot initialization. A PPO-based RL stage then adapts only the per-joint input scaling factors for deployment-specific refinement. The paper reports up to 80.4% error reduction on a challenging Franka joint (12.36° to 2.42°), 19.2% improvement under parameter uncertainty, and an "optimization ceiling effect" in which RL helps only when the meta-learned baseline has localized high-error joints. The work is simulation-only, with a proposed staged deployment protocol for future real-world validation.

Significance. If the central claims were fully supported, the framework would be a practical contribution: it offers a physically grounded way to initialize an interpretable fuzzy-PID controller from robot features, a compact RL adaptation space that avoids catastrophic forgetting, and clear guidance on when RL refinement is worthwhile. The paper also contains genuine strengths: a detailed reproducibility appendix, fixed training seeds, a 100-seed robustness evaluation protocol, distribution-shift checks on the augmented dataset, and an honest statement that real-world deployment remains untested. However, the current experimental design does not actually evaluate cross-platform generalization, because the two test platforms are also base training platforms. In addition, the baseline methods defined in Section 4.5 are never reported in the results, and several load-bearing numerical details are internally inconsistent. These issues make the headline claims unsupported as written, though they appear addressable within the manuscript's simulation-based scope.

major comments (5)
  1. [§4.2.1, §4.4.1; Tables 2–3] The claimed cross-platform generalization is not tested. Franka Panda and Laikago are two of the three base robots used to generate the meta-training set (§4.2.1, Figure 2), yet §4.4.1 evaluates zero-shot Meta-LF-PID and Meta-LF-PID+RL on exactly these platforms. The statement that they are "seen during RL training (only used in meta-learning)" is misleading because the meta-network was trained on virtual variants of both platforms; the results are in-distribution interpolation, not transfer to a new morphology. The RL results on Franka are also obtained after training RL on Franka itself, so they show per-platform fine-tuning, not cross-platform adaptation. I request evaluation on at least one held-out morphology not represented among the three bases (e.g., a 6-DOF arm or a biped), with the same protocol. Without such a test, the "unseen platform" language and the abstract's "cross-plat
  2. [§4.5, Tables 2–5] The four baseline families defined in §4.5 — manual Fuzzy-PID, Ziegler–Nichols, DE-optimized LF-PID, and pure Meta-LF-PID — are never reported in the results. All tables and figures in Section 5 compare only Meta-LF-PID vs. Meta-LF-PID+RL. Consequently, the paper does not support claims such as "improves... over classical methods" (Abstract) or the efficiency advantage argued in §4.5 and Appendix C. The key comparative questions listed in §4.5 remain unanswered. Please provide numerical results for all baselines under identical evaluation conditions (same trajectories, disturbance scenarios, and seeds), or explicitly restrict the paper's claims to the comparison between meta-learning and meta-learning-plus-RL.
  3. [§3.2.1, Figure 1, §4.2.2, Appendix A.1] There are irreconcilable inconsistencies in the description of the meta-learning architecture and dataset. Equation (5) gives D_total = 6n(1+2R) = 330n, so Franka's output should be 2970 and Laikago's 3960; §3.2.1 repeats this. Yet Figure 1 states "Output: 3×7=21" and reports 104,789 total parameters, and the same figure's training statistics say "Samples: 303," while the caption and §4.2.2 say the network is trained on N=232 filtered variants. §4.3 says meta-training takes 5 minutes / 500 epochs, while Appendix A.1 says about 8 minutes / 500 epochs with early stopping. These discrepancies make the architecture and the overfitting claims (Appendix E) impossible to verify. Please align the text, equation, figure, and appendix, and state clearly which output dimension and sample count were actually used.
  4. [§3.2.2, §3.3.2, Appendix A.4, Appendix A.5] Several load-bearing training and evaluation parameters are inconsistent. RL training time is reported as "~10 minutes" in §3.2.2, §4.3, and §5.3, but Appendix A.5, Table 11 gives "RL (per robot) ~2.5 hours." Control frequency is given as 240 Hz in §3.3.2 (Eq. 17 context), §3.4, and §4.4.1, but Appendix A.4, Table 10 lists "Control frequency 100 Hz" with a physics timestep of 0.01 s. The reward function in Eq. (16) uses weights α1=10.0, α2=0.1, α3=0.1, but Appendix A.4, Table 10 defines a different reward with weights 1.0, 0.5, 0.1, 0.05 and explicit success/failure terms. These inconsistencies affect the reported efficiency, the 10,000-episode robustness total, and the reproducibility claims. Please provide a single consistent configuration, or explain which numbers apply to which experiments.
  5. [§6.2.3] The "optimization ceiling effect" is presented as a general design principle, with the quantitative guidance that "heterogeneous profiles (CV>0.4) benefit maximally, while uniform profiles (CV<0.2) achieve minimal gains." However, this is derived from only two test platforms, both of which are also training bases. No table or scatter plot of CV vs. RL improvement is provided, and no statistical test supports the specific CV thresholds. With n=2, this is an anecdotal observation, not a validated finding. Either provide data across additional held-out morphologies to substantiate the thresholds, or explicitly present the ceiling effect as a hypothesis for future work.
minor comments (5)
  1. [§4.2.2] The dataset counts are unclear: the text says 150 variants are generated for Franka and 150 for KUKA, with "Base robots: 3," but no Laikago variants are mentioned. The total 303 = 150 + 150 + 3 suggests Laikago appears only as a base, which conflicts with Algorithm 1's loop over all three bases and with the claim of 232 samples from "three base platforms." Please clarify which base generated which variants.
  2. [Appendix D.2] There is a formatting artifact in the text: "extbfLaikago trend (12-DOF)" should read "Laikago trend (12-DOF)."
  3. [References [37] and [39]] References [37] and [39] are duplicate entries for the same publication (OpenAI et al., "Learning dexterous in-hand manipulation"). Please merge them.
  4. [§3.2.1, Eq. (11)] The output denormalization after the sigmoid is described only as "element-wise affine transformations calibrated to typical PID gain ranges and consequent magnitudes." Please give the exact de-normalization ranges for base gains, scaling factors, and consequent parameters; these are needed to interpret Eq. (11) and the optimization bounds in Appendix A.3.
  5. [§4.4.1 and Abstract] The phrases "unseen platform" and "cross-platform generalization on two distinct systems" overstate the design. Since Franka and Laikago are base training robots, the evaluation should be described as within-distribution zero-shot prediction plus platform-specific RL refinement, unless a true held-out morphology is added.

Circularity Check

1 steps flagged · score 6.0 of 10

Test platforms are meta-training base robots, so the zero-shot 'cross-platform' results are an in-distribution fit check; the RL-vs-baseline comparisons remain genuine optimization results.

  1. fitted input called prediction [§4.2.1 (Base Robots) and §4.4.1 (Cross-Platform Generalization)]
    "We use 3 base robot platforms as training data sources for physics-based data augmentation, as shown in Figure 2: Franka Panda (9-DOF manipulator), KUKA LBR iiwa (7-DOF redundant manipulator), and Laikago (12-DOF quadruped). ... For each base robot, we perform careful LF-PID optimization using the hybrid differential evolution and Nelder-Mead strategy (Algorithm 2) to obtain ground-truth optimal parameters. This provides high-quality supervision for the meta-learning stage. ... We evaluate on both Franka Panda and Laikago platforms, neither of which is seen during RL training (only used in met"

    Franka and Laikago are not external test platforms: they are two of the three base robots whose optimized LF-PID parameters are used as meta-learning supervision (Eq. 12), and their variants form the 232-sample training set. The zero-shot Meta-LF-PID evaluation on them therefore measures how well the network reproduces parameters it was trained to fit for those morphologies, not how well it transfers to an unseen morphology. The sentence 'only used in meta-learning' admits this overlap while still calling the evaluation cross-platform. Consequently the headline 'cross-platform generalization' result reduces to an in-distribution fit/interpolation check; the only genuinely optimization-based claim (RL on top of the meta baseline) remains independent.

full rationale

The only load-bearing circularity is in the evaluation design: Franka and Laikago are two of the three base robots whose optimized parameters supervise the meta-network (Eq. 12), yet §4.4.1 presents them as cross-platform test cases. The zero-shot Meta-LF-PID numbers are therefore a validation of the network's ability to reproduce training-morphology parameters, not a test of transfer to a new robot; the phrase 'only used in meta-learning' concedes this overlap. The RL-adaptation results are not circular in the same sense: they are genuine within-simulator optimization comparisons against a fixed meta baseline, and the LF-PID formulation itself is not defined in terms of the evaluation metric. There is no self-citation chain, no imported uniqueness theorem, and no ansatz smuggled in by citation. The 'optimization ceiling effect' is close to a restatement of the percentage-improvement arithmetic, but it is a minor interpretive label rather than a load-bearing derivation. Score 6 reflects the partial collapse of the central cross-platform prediction claim into a fit evaluation.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

No new physical entities are introduced; the LF-PID parameterization and the 'optimization ceiling effect' are method/insight constructs, not entities with independent falsifiable handles. The central claim rests on hand-set perturbation ranges, filter thresholds, reward weights, and unverified optimizer convergence.

free parameters (7)
  • Perturbation ranges for virtual robot synthesis = mass ±10%, inertia ±15%, friction ±20%, length ±5%
    Hand-selected ranges define the training distribution and are presented as the paper's physics-constrained augmentation; no tuning or justification beyond cited manufacturing tolerances.
  • Controllability filter threshold = 30°
    Samples with optimized LF-PID tracking error >30° are discarded (71/303). This loose threshold selects optimization success, not physical validity, and shapes the meta-learning labels.
  • Reward weights α1, α2, α3 = 10.0, 0.1, 0.1
    Hand-selected; Appendix A.4 lists a different reward (position weight 1.0, velocity 0.5, jerk 0.1, parameter-change 0.05), so it is unclear which reward produced the reported RL results.
  • RL adaptation bounds = Δs ∈ [−0.2, 0.2]; s clipped to [0.1, 10.0]
    Hand-set action limits determine how much the RL stage can change scaling factors; central to the 16.6% improvement claim.
  • Fuzzy membership partitions = not specified
    Shared Gaussian centers/widths are declared fixed ('membership partitions are shared and fixed') but their values are never given; they determine the gain-scheduling behavior.
  • DE and Nelder-Mead bounds = Kp,Kd ∈ [0.1,500]; Ki ∈ [0,1]
    Hand-set optimizer bounds constrain the ground-truth LF-PID labels that supervise meta-learning.
  • Output denormalization calibration = not specified
    The sigmoid network outputs [0,1] are mapped to control gains via 'affine transformations calibrated to typical PID gain ranges'; the calibration constants are not reported.
assumptions (5)
  • domain assumption PyBullet position-controlled simulation with sinusoidal reference trajectories is a faithful proxy for real robot tracking performance.
    All quantitative claims (80.4%, 19.2%, 16.6%) are measured in PyBullet only; Section 6.1.1 concedes real-world factors (sensor noise, unmodeled friction, latency) are untested.
  • domain assumption The 10D feature vector (DOF, mass, inertia, link lengths, COM, friction) is sufficient to predict near-optimal LF-PID parameters across morphologies.
    Meta-learning maps f∈R^10 to θ of dimension 330n; Eq. (7) in §3.2.1. If features omit control-relevant dynamics, cross-platform initialization fails.
  • domain assumption Hybrid DE+Nelder-Mead with 120 evaluations per virtual robot finds near-global optimal LF-PID parameters.
    These optima are the supervised labels θ*_v for meta-learning (Algorithm 2, §3.3.2). With population 8 and 15 DE iterations plus 20 NM iterations, convergence is not verified.
  • ad hoc to paper Tracking error <30° is a meaningful controllability/quality threshold.
    Used to filter 71/303 training variants (§3.3.1); 30° is a very loose tracking-error bound for manipulators, so the 'physically valid' dataset is partly determined by this arbitrary threshold.
  • domain assumption Randomized sinusoidal trajectories cover deployment conditions.
    Reference trajectories are sinusoids with randomized amplitude/frequency/phase (§3.4 Eq. 20); no contact-rich or task-specific trajectories are tested.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Cross-Platform Learnable Fuzzy Gain-Scheduled Proportional-Integral-Derivative Controller Tuning via Physics-Constrained Meta-Learning and Reinforcement Learning Adaptation." pith.science (2026). https://pith.science/paper/5IGTQ2MF

@misc{pith2026251106500,
  author       = {Pith},
  title        = {Pith review of: Cross-Platform Learnable Fuzzy Gain-Scheduled Proportional-Integral-Derivative Controller Tuning via Physics-Constrained Meta-Learning and Reinforcement Learning Adaptation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5IGTQ2MF}},
  note         = {Machine review of arXiv:2511.06500}
}
read the original abstract

Motivation and gap: PID-family controllers remain a pragmatic choice for many robotic systems due to their simplicity and interpretability, but tuning stable, high-performing gains is time-consuming and typically non-transferable across robot morphologies, payloads, and deployment conditions. Fuzzy gain scheduling can provide interpretable online adjustment, yet its per-joint scaling and consequent parameters are platform-dependent and difficult to tune systematically. Proposed approach: We propose a hierarchical framework for cross-platform tuning of a learnable fuzzy gain-scheduled PID (LF-PID). The controller uses shared fuzzy membership partitions to preserve common error semantics, while learning per-joint scaling and Takagi-Sugeno consequent parameters that schedule PID gains online. Combined with physics-constrained virtual robot synthesis, meta-learning provides cross-platform initialization from robot physical features, and a lightweight reinforcement learning (RL) stage performs deployment-specific refinement under dynamics mismatch. Starting from three base simulated platforms, we generate 232 physically valid training variants via bounded perturbations of mass (+/-10%), inertia (+/-15%), and friction (+/-20%). Results and insight: We evaluate cross-platform generalization on two distinct systems (a 9-DOF serial manipulator and a 12-DOF quadruped) under multiple disturbance scenarios. The RL adaptation stage improves tracking performance on top of the meta-initialized controller, with up to 80.4% error reduction in challenging high-load joints (12.36 degrees to 2.42 degrees) and 19.2% improvement under parameter uncertainty. We further identify an optimization ceiling effect: online refinement yields substantial gains when the meta-initialized baseline exhibits localized deficiencies, but provides limited improvement when baseline quality is already uniformly strong.

Figures

Figures reproduced from arXiv: 2511.06500 by the authors.

Figure 1
Figure 1. Meta-LF-PID Network Architecture. The hierarchical feedforward network consists of an input layer (10D robot features 𝐟 including mass, DOF, inertia, link lengths, and friction), two encoder layers with LayerNorm and ReLU activations (256D: 𝐡1 , 𝐡2 ), a hidden layer (128D: 𝐡ℎ𝑖𝑑𝑑𝑒𝑛), and structured parameter heads that output bounded LF-PID initialization 𝜽̂ ∈ [0, 1]𝐷 via sigmoid activation (𝜎). Here 𝜽̂ contains only… view at source ↗
Figure 2
Figure 2. Three base robot platforms serving as training data sources for physics-based data augmentation in PyBullet simulation. (a) Franka Panda manipulator (9-DOF) with complex serial kinematics, (b) KUKA LBR iiwa redundant manipulator (7-DOF) offering enhanced dexterity, and (c) Laikago quadruped (12-DOF) with parallel leg structure. From these diverse platforms, we generate 303 virtual variants through systematic perturb… view at source ↗
Figure 3
Figure 3. Cross-platform generalization: Per-joint tracking error comparison across two morphologically distinct robot platforms. (a) Franka Panda serial manipulator (9-DOF) achieves 16.6% overall improvement with exceptional gains in high-load joints (J2: +80.4%, from 12.36° to 2.42°), demonstrating highly effective adaptation to manipulation tasks with concentrated loads. (b) Laikago parallel quadruped (12-DOF) achieves 2.1… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Comprehensive tracking performance comparison on Franka Panda. (a) Actual tracking error time series showing 10.9% improvement with RL adaptation—RL reduces tracking oscillations and peak errors by smoothing control responses, (b) Error distribution histograms demonstr…
Figure 5
Figure 5. Figure 5: Robustness evaluation across five disturbance scenarios on Franka Panda (20 episodes per scenario, evaluated across 100 random seeds for stochastic validation). Subplots (a-c) show detailed results for representative seed 51 (near-median performance), while subplot (d)…
Figure 6
Figure 6. Figure 6: Comprehensive RL training dynamics monitoring dashboard for Franka Panda (9-DOF) over 1M timesteps using PPO algorithm with optimized hyperparameters. (a) Episode reward improves progressively, demonstrating effective learning. (b) Value function loss decreases logarit…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

39 extracted references · 6 linked inside Pith

  1. [1]

    K. J. Åström, T. Hägglund, Advanced PID Control, ISA-The Instru- mentation, Systems, and Automation Society, 2006

    2006

  2. [2]

    R.Vilanova,A.Visioli,PIDControlintheThirdMillennium:Lessons Learned and New Approaches, Springer-Verlag, London, UK, 2012

    2012

  3. [3]

    C. D. Johnson, M. A. Franchek, Real-world challenges in industrial robot controller tuning: A case study analysis, IEEE Trans. Autom. Sci. Eng. 18 (3) (2021) 1456–1468

    2021

  4. [4]

    Gaing, A particle swarm optimization approach for optimum designofPIDcontrollerinAVRsystem,IEEETrans.EnergyConvers

    Z.-L. Gaing, A particle swarm optimization approach for optimum designofPIDcontrollerinAVRsystem,IEEETrans.EnergyConvers. 19 (2) (2004) 384–391

    2004

  5. [5]

    Berkenkamp, A

    F. Berkenkamp, A. P. Schoellig, A. Krause, Safe controller optimiza- tion for quadrotors with Gaussian processes, in: Proc. IEEE ICRA, 2016, pp. 491–496

    2016

  6. [6]

    T. P. Lillicrap, J. J. Hunt, A. Pritzel, N. Heess, T. Erez, Y. Tassa, D. Silver, D. Wierstra, Continuous control with deep reinforcement learning, arXiv preprint arXiv:1509.02971 (2015)

    arXiv 2015

  7. [7]

    Zhang, B

    Y. Zhang, B. Nie, Z. Cao, Y. Fu, Y. Gao, Disturbance-aware adaptive compensation in hybrid force-position locomotion policy for legged robots, in: Proc. IEEE ICRA, 2024, arXiv:2506.00472

    arXiv 2024

  8. [8]

    I.C.Trelea,Theparticleswarmoptimizationalgorithm:convergence analysis and parameter selection, Inf. Process. Lett. 85 (6) (2003) 317–325. Wu et al.:Preprint submitted to ElsevierPage 23 of 24 Cross-Platform LF-PID Tuning via Meta-RL

    2003

Show all 39 references
  1. [9]

    Schulman, F

    J. Schulman, F. Wolski, P. Dhariwal, A. Radford, O. Klimov, Proxi- malpolicyoptimizationalgorithms,arXivpreprintarXiv:1707.06347 (2017)

    2017 arXiv

  2. [10]

    Nagabandi, G

    A. Nagabandi, G. Kahn, R. S. Fearing, S. Levine, Neural network dynamics for model-based deep reinforcement learning with model- free fine-tuning, in: Proc. IEEE ICRA, 2018, pp. 7559–7566

    2018

  3. [11]

    X. Yu, P. He, Z. Wan, A. Tsukada, An adaptive SAC-PID control method based on reinforcement learning for mobile robots, in: Proc. IEEE ICRA, 2021, arXiv:2103.10686

    2021 arXiv

  4. [12]

    Jiang, Z

    D. Jiang, Z. Li, Y. Xia, Reinforcement learning based adaptive track- ing control for continuum robots, J. System Simulation 34 (7) (2022) 1465–1475. (in Chinese)

    2022

  5. [13]

    Pezzato, R

    C. Pezzato, R. Ferrari, C. Hernandez Corbato, A novel adaptive controller for robot manipulators based on active inference, IEEE Robotics Autom. Lett. 5 (2) (2020) 2973–2980

    2020

  6. [14]

    Hospedales, A

    T. Hospedales, A. Antoniou, P. Micaelli, A. Storkey, Meta-learning inneuralnetworks:Asurvey,IEEETrans.PatternAnal.Mach.Intell. 44 (9) (2022) 5149–5169

    2022

  7. [15]

    C. Finn, P. Abbeel, S. Levine, Model-agnostic meta-learning for fast adaptation of deep networks, in: Proc. ICML, 2017, pp. 1126–1135

    2017

  8. [16]

    C. Finn, T. Yu, T. Zhang, P. Abbeel, S. Levine, One-shot visual imitation learning via meta-learning, in: Proc. CoRL, 2017, pp. 357– 368

    2017

  9. [17]

    W. Yu, C. K. Liu, G. Turk, Meta-world: A benchmark and evaluation formulti-taskandmetareinforcementlearning,in:Proc.CoRL,2020, pp. 1094–1100

    2020

  10. [18]

    IEEE CDC, 2024, arXiv:2410.07575

    W.He,Y.Hu,Y.Guan,M.Xue,Self-supervisedmeta-learningforall- layer DNN-based adaptive control with stability guarantees, in: Proc. IEEE CDC, 2024, arXiv:2410.07575

    2024 arXiv

  11. [19]

    Tobin, R

    J. Tobin, R. Fong, A. Ray, J. Schneider, W. Zaremba, P. Abbeel, Domain randomization for transferring deep neural networks from simulationtotherealworld,in:Proc.IEEE/RSJIROS,2017,pp.23– 30

    2017

  12. [20]

    X. B. Peng, M. Andrychowicz, W. Zaremba, P. Abbeel, Sim-to-real transfer of robotic control with dynamics randomization, in: Proc. IEEE ICRA, 2018, pp. 3803–3810

    2018

  13. [21]

    Kumar, Z

    A. Kumar, Z. Fu, D. Pathak, J. Malik, RMA: Rapid motor adaptation for legged robots, in: Proc. CoRL, 2021, pp. 1034–1045

    2021

  14. [22]

    Okamoto, S

    S. Okamoto, S. Nagano, M. Kojima, Robust fault-tolerant control of quadrupedrobotsusingadaptivecurriculumdynamicrandomization, IEEE Access 9 (2021) 150385–150396

    2021

  15. [23]

    Shorten, T

    C. Shorten, T. M. Khoshgoftaar, A survey on image data augmenta- tion for deep learning, J. Big Data 6 (1) (2019) 1–48

    2019

  16. [24]

    Todorov, T

    E. Todorov, T. Erez, Y. Tassa, MuJoCo: A physics engine for model- based control, in: Proc. IEEE/RSJ IROS, 2012, pp. 5026–5033

    2012

  17. [25]

    Storn, K

    R. Storn, K. Price, Differential evolution – A simple and efficient heuristic for global optimization over continuous spaces, J. Global Optim. 11 (4) (1997) 341–359

    1997

  18. [26]

    J. A. Nelder, R. Mead, A simplex method for function minimization, Comput. J. 7 (4) (1965) 308–313

    1965

  19. [27]

    [Online]

    E.Coumans,Y.Bai,PyBullet,aPythonmoduleforphysicssimulation for games, robotics and machine learning, 2016. [Online]. Available: http://pybullet.org

    2016

  20. [28]

    International Organization for Standardization, ISO 9283:1998 Ma- nipulating Industrial Robots—Performance Criteria and Related Test Methods, ISO Standards, Geneva, Switzerland, 1998

    1998

  21. [29]

    J. Cho, S. Kim, Parameter identification and model validation of in- dustrialrobotmanipulatorsundermanufacturingtolerances,Robotica 37 (10) (2019) 1656–1672

    2019

  22. [30]

    S. Lee, H. Park, J. Kim, Real-world parameter uncertainty quantifi- cationincollaborativerobots:Anexperimentalstudy,IEEERobotics Autom. Lett. 7 (4) (2022) 9847–9854

    2022

  23. [31]

    Collins, S

    J. Collins, S. Chand, A. Vanderkop, D. Howard, A review of physics simulators for robotic applications, IEEE Access 9 (2021) 51416– 51431

    2021

  24. [32]

    J. Tan, T. Zhang, E. Coumans, et al., Sim-to-real: Learning agile locomotion for quadruped robots, in: Proc. Robotics: Science and Systems (RSS), 2018

    2018

  25. [33]

    G. B. Margolis, T. Yang, K. Paigwar, X. Chen, P. Agrawal, Rapid locomotion via reinforcement learning, in: Proc. Robotics: Science and Systems (RSS), 2024, arXiv:2404.01063

    2024 arXiv

  26. [34]

    Franka Emika GmbH, Franka Panda Performance Benchmarks: Po- sition Control Accuracy in Industrial Settings, Technical Report FE- 2021-03, Munich, Germany, 2021

    2021

  27. [35]

    C. Ott, A. Dietrich, A. Albu-Schäffer, Unified impedance and admit- tance control, in: Proc. IEEE ICRA, 2017, pp. 554–560

    2017

  28. [36]

    W. Zhao, J. P. Queralta, T. Westerlund, Sim-to-real transfer in deep reinforcement learning for robotics: A survey, in: Proc. IEEE SSCI, 2020, pp. 737–744

    2020

  29. [37]

    OpenAI et al., Learning dexterous in-hand manipulation, Int. J. Robotics Res. 39 (1) (2020) 3–20

    2020

  30. [38]

    Berkenkamp, A

    F. Berkenkamp, A. P. Schoellig, Safe exploration in reinforcement learning: Theory and applications in robotics, in: Proc. European Control Conference (ECC), 2021, pp. 1–19

    2021

  31. [39]

    Andrychowicz, B

    M. Andrychowicz, B. Baker, M. Chociej, et al., Learning dexterous in-hand manipulation, Int. J. Robotics Res. 39 (1) (2020) 3–20. Wu et al.:Preprint submitted to ElsevierPage 24 of 24

    2020

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.