Pith. sign in

REVIEW 6 major objections 6 minor 46 references

SPI-BoTER: Error Compensation for Industrial Robots via Sparse Attention Masking and Hybrid Loss with Spatial-Physical Information

T0 review · 6 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read SPI-BoTER combines a Denavit-Hartenberg kinematics branch with a sparse-attention Transformer and a distance-matrix physics loss to reduce UR5 3D positioning error to 0.2515 mm, 35.16% below a standard DNN.

desk verdict A plausible engineering paper whose central physics-loss claim is not demonstrated because the adaptive weighting, as described, collapses to zero and several reported numbers are internally inconsistent. read the letter →

arxiv 2506.22788 v1 pith:JRSX2OER submitted 2025-06-28 cs.RO

Xuao Hou , Yongquan Jia , Shijin Zhang , Yuqiang Wu This is my paper
classification cs.RO
keywords Mechanism-datafusionErrorcompensationRoboticarmposepredictionPhysics-informedneuralnetworksTransformerSparseattentionmaskingHybridlossUR5robot
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

The paper proposes SPI-BoTER, a dual-branch network for industrial robot error compensation that combines a physics-based Denavit-Hartenberg forward kinematics branch with a Transformer branch that predicts residual errors. The Transformer uses a sparse self-attention mask shaped by the six-axis kinematic chain, and training is driven by a hybrid loss that adds a spatial-physics term aligning normalized Euclidean distance matrices of predicted and theoretical positions. The method is designed to work from small datasets: on 724 UR5 samples it reports a 3D mean absolute error of 0.2515 mm, a 35.16% reduction over a standard DNN baseline, and an inverse joint-angle compensation routine that converges to 0.01 mm in about 147 iterations. The paper's contribution is showing that kinematic structure and spatial topology can be injected into a Transformer to reach sub-millimeter accuracy with limited data.

What carries the argument

The two load-bearing mechanisms are the sparse self-attention mask and the spatial-physics loss. The mask is a 6x6 binary matrix built from the kinematic chain's functional grouping: joints 1-3 (coarse positioning) are fully mutually visible, joints 4-6 (fine orientation) are fully mutually visible, and joints 1-3 may attend to joints 4-6 but not conversely; masked entries are set to a large value added to attention scores. The loss term L_physics compares normalized Euclidean distance matrices D_pred-norm and D_theory-norm across a training batch, enforcing that predicted points preserve the theoretical model's relative spatial topology. The paper's claim is that these two components, together with a trainable fusion scale alpha initialized to 0.1, let the network learn physically consistent residual errors from only 724 samples.

What would settle it

Track the learned lambda_data and lambda_physics during a training run; if either goes to zero while total loss collapses, the reported ablation gain cannot be attributed to the SPI loss. Compare against a fixed-positive-weight version of the same loss.

Watch

Extended reading notes

Core claim

The paper claims that error compensation for a six-axis industrial robot can be recast as two coupled problems: a forward prediction task in which a dual-branch network predicts end-effector position as a theoretical DH-model position plus a learned residual, and an inverse task in which the frozen network is used as a differentiable forward solver to find joint-angle corrections by gradient descent. The forward branch is a Transformer encoder whose attention is constrained by a sparse mask that groups the first three joints (coarse position) and last three joints (fine orientation), with unidirectional visibility from the first group to the second. Training uses a hybrid loss that combines a data residual with a spatial-physics residual computed as the mean squared difference between normalized Euclidean distance matrices of predicted and theoretical positions. On a 724-sample UR5 dataset split 8:1:1, the paper reports a 3D mean absolute error of 0.2515 mm, a 35.16% reduction over a standard DNN, and an inverse compensation routine converging to 0.01 mm accuracy in about 147 iterations.

Load-bearing premise

The adaptive loss weights lambda_data and lambda_physics are trained by gradients equal to the loss values themselves, which are never negative, so nothing in the described update prevents the weights from collapsing to zero and switching off both the data fit and the physics constraint.

Editorial extensions

If this is right

  • If the reported result holds, a sub-millimeter (0.25 mm) absolute positioning error is attainable from 724 static samples on a UR5-class arm, which is in the range needed for precision cutting and welding operations.
  • The inverse compensation routine's 147-iteration convergence to 0.01 mm suggests the trained forward model can serve as a differentiable solver for closed-loop joint-angle correction without retraining.
  • The mask's reduction in required epochs (111 vs 402 without mask, per Table 8) implies that kinematic structure can be injected cheaply as an attention prior, improving training efficiency as well as accuracy.
  • Because the ablation shows a 12% drop when the physics loss is removed, the spatial distance-matrix constraint is presented as the component suppressing outlier errors.
  • The framework's stated limitation is that it is validated on static randomly sampled points, not continuous trajectories, so trajectory-level performance is an open next step.

Reading between the lines

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

  • Beyond the paper: the same dual-branch design could be re-derived for other serial arms by replacing the UR5-specific mask with any chain-consistent grouping, and whether the 35% gain survives on a different kinematic structure is a direct testable extension.
  • Beyond the paper: the distance-matrix loss normalizes by each batch's maximum distance, so its gradient depends on batch composition; with very small batches this could make the physics constraint noisy, a property the paper does not analyze.
  • Beyond the paper: if the adaptive weights collapse toward zero as their gradients suggest, the reported ablation gains may be attributable to the sparse mask and residual head rather than the SPI loss, and tracking the learned lambdas would separate these effects.
  • Beyond the paper: because the physics branch inherits the nominal DH model's errors, coupling the same loss with an identified kinematic model rather than the published DH table is a natural next step.
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

6 major / 6 minor

Summary. The manuscript presents SPI-BoTER, a dual-branch architecture for six-axis robot end-effector error compensation: a differentiable Denavit-Hartenberg forward-kinematics branch, a Transformer branch with a hand-designed sparse self-attention mask and a residual prediction head, and a hybrid loss that combines a data MSE term with a normalized Euclidean distance-matrix mismatch between predicted point sets and theoretical DH-model point sets. The loss weights are claimed to be adaptively learned. A gradient-descent inverse compensation procedure is then applied to the trained model. On a 724-sample UR5 dataset split 8:1:1, the authors report a 3D MAE of 0.2515 mm, a 35.16% reduction over their DNN baseline, and convergence of the inverse compensation to 0.01 mm within 147 iterations. The paper also reports ablations over loss terms and architectural components.

Significance. If substantiated, the work would be a useful empirical contribution to robot absolute-accuracy compensation, particularly in the small-sample regime; the idea of structuring attention by manipulator kinematics and using distance-matrix consistency as a soft regularizer is plausible and worth pursuing. The paper deserves credit for using real measurement data from a UR5/TrackScan setup, for evaluating on a physical platform, and for explicitly listing limitations in the conclusion. However, the significance is currently undermined by several load-bearing technical inconsistencies: the internal inconsistency of Table 3, the incorrect mask-addition operation in Eq. (6), the collapse behavior of the adaptive weights in Eqs. (17)-(18), and the contradiction between the text and Table 9. These issues prevent the results from being interpreted as stated.

major comments (6)
  1. [Table 3 / Section 4.2.1] The main results table is internally inconsistent. In the first block, the MAE and MSE columns are numerically identical for every axis (X: 0.0190/0.0190, Y: 0.0422/0.0422, Z: 0.0175/0.0175), which is impossible for nonzero errors because MSE is a squared quantity and can equal MAE only in degenerate cases; moreover, with the reported RMSE values around 0.13-0.21, the corresponding MSE values would be about 0.02-0.04, not 0.019-0.042. In the third block the 3D row reads MAE=0.2515, MSE=0.2850, RMSE=0.0812, while Table 4 for the same model reports MAE=0.2515, MSE=0.0812, RMSE=0.2850. The table also labels all three blocks 'Train 580' although the surrounding text describes train, validation, and test sets. Because the headline 0.2515 mm value and the R2 values used in Tables 4, 6, and 8 are drawn from this block, a corrected and verified table is required before the central numerical claims can be assessed.
  2. [Section 2.1.2, Eq. (4)-(6)] The sparse mask is implemented in a way that would have the opposite of the intended effect. Equation (4) defines M with 1 = masked and 0 = visible, and Eq. (6) computes MaskedAttenScores = AttenScores + M. Adding +1 to masked positions makes those positions numerically larger before the softmax, i.e., more attended, while visible positions are unchanged; standard masking would add a large negative number (for example, -1e9) or use a boolean key_padding_mask. As described, the sparse attention mask cannot explain the improvements attributed to it in Section 4.2.4, and the reported differences between the No Mask, Body Mask, and Baseline groups would need a different explanation.
  3. [Section 2.2.3, Eqs. (17)-(18)] The adaptive weighting scheme as described collapses. The trainable weights λ_data and λ_physics, or their log-parameterized forms, receive gradients equal to the corresponding nonnegative losses L_data and L_physics. Unconstrained gradient descent therefore drives the weights toward zero, making L_total tend to zero and removing both data fitting and the physics term; the exp(log λ) reparameterization in Eq. (17) does not provide any lower bound. No normalization, sum-to-one constraint, regularization, or uncertainty-based weighting is described, and no learned λ values or training curves are reported. Without such evidence, the 12% gap between the Baseline and Data Only groups in Table 6 cannot be attributed to the spatial-physics loss.
  4. [Section 2.2.2 and Eq. (9)] The 'physics' loss is not an independent physical law. Because p_pred = p_theory + α·Δp by Eq. (9), the distance-matrix fidelity term in Eq. (16) aligns D(p_pred) with D(p_theory) and therefore partly just penalizes large or structurally distorted Δp relative to the DH model. This may be a useful regularizer, but the abstract and contribution 2 describe the method as enforcing 'physical consistency' and 'physical interpretability'; as presented, the constraint is better characterized as a soft prior toward the theoretical kinematic model.
  5. [Section 5, Table 9] The inverse-compensation results are internally contradictory. The text states that the maximum absolute deviation was 0.62 mm and the maximum standard deviation was 0.1495 mm, but Table 9 lists per-axis maximum errors of 0.4850, 0.7240, and 0.8220 mm, whose maximum is 0.8220 mm, and per-axis standard deviations whose maximum is 0.1438 mm. The abstract's statement that the inverse algorithm 'converges to an accuracy of 0.01 mm' also conflates the stopping threshold (MSE below 1e-4 in Section 2.3.1) with achieved positioning accuracy, and Table 9 reports minimum errors as large as 0.026 mm. These numbers need to be reconciled and the accuracy claim reworded.
  6. [Section 3.2 and Section 4.2] All headline comparisons rest on a single random split (seed 7) and a single training run (seed 139), with no repeated runs or error bars. On a test set of roughly 72 samples, the reported 35.16% MAE reduction relative to DNN and the 12% ablation gap in Table 6 could be within run-to-run variability. The authors should report mean and standard deviation over at least five to ten random seeds, or provide a statistical significance test, before claiming a reliable improvement.
minor comments (6)
  1. [Section 4.2.1] The text refers to 'Table ??' instead of Table 3, and the table's first column labels for the three blocks appear to have been lost.
  2. [Section 4.1, Eqs. (39)-(42)] The R2 formulas use Δx_r, described as the arithmetic mean of the absolute prediction errors, in the denominator. The standard definition of R2 uses the variance of the actual target values around their mean; as written, Eqs. (39)-(42) do not compute the coefficient of determination.
  3. [Section 2.3.1] The four enumerated steps in the inverse algorithm are all labeled 'Step 1'; they should be numbered Steps 1-4 for clarity.
  4. [Section 3.2] The text says the model was trained for 150 epochs and achieved its best performance at epoch 111, but the hyperparameter bullet later states a maximum of 5000 training epochs; these numbers should be reconciled.
  5. [Section 6] The conclusion claims a 34% parameter-count reduction and 97.3% spatial similarity, but neither quantity is defined or reported in the experimental section; please add supporting definitions and results.
  6. [Section 2.1.2] The subsection contains two consecutive headings numbered '(3)', which makes the structure of the output-layer description confusing.

Circularity Check

3 steps flagged · score 4.0 of 10

Central forward MAE result is self-contained against measured data; the by-construction items are the 0.01 mm inverse-convergence claim (the early-stopping threshold), the 97.3% 'spatial similarity' (a restatement of the SPI loss), and a self-defeating adaptive lambda update that can switch off the physics term.

  1. self definitional [Section 2.3.1, Early Stopping; Abstract]
    "If the MSE loss drops below a threshold of 1×10−4, terminate the iteration early to accelerate computation. ... the inverse angle compensation algorithm converges to an accuracy of 0.01 mm within an average of 147 iterations."

    The reported 'accuracy of 0.01 mm' is exactly sqrt(1e-4), i.e. the RMSE corresponding to the MSE early-stopping threshold. The optimization is stopped when the model's own forward residual reaches this value, so the convergence accuracy is fixed by the stopping rule by construction. Only the average iteration count of 147 is an empirical observation. The independent physical validation in Table 9 reports much larger actual errors (max 0.62 mm), so this particular circular metric is not load-bearing for the main forward-error-compensation claim.

  2. self definitional [Section 6 Conclusion; Section 2.2.2, Eq. (16)]
    "The hybrid spatial-physics loss function aligns predictions with the topological structure of the theoretical model through a Euclidean distance matrix alignment mechanism, increasing spatial similarity to 97.3% and effectively suppressing extreme errors under challenging working conditions."

    Equation (16) defines Lphysics as the MSE between the normalized predicted and theoretical Euclidean distance matrices. The '97.3% spatial similarity' reported in the conclusion is therefore a monotone rescaling of the very objective that was minimized during training. Claiming it as an achieved property is a restatement of the loss definition rather than an independent physical validation. This is a secondary descriptive claim and does not by itself invalidate the forward position-prediction result.

1 more flagged steps
  1. other [Section 2.2.3, Eqs. (17)-(18); ablation Table 6]
    "To prevent the weights from taking negative values, logarithmic parameterization is adopted: λ∗ = exp(logλ∗). ... Ltotal = λdataLdata + λphysicsLphysics ... these weight parameters are automatically updated through backpropagation."

    The gradients of Ltotal with respect to λdata and λphysics are exactly Ldata and Lphysics, both non-negative MSE terms. Because the paper describes no normalization, no positivity-preserving lower bound beyond exp(s)>0, and no reported learned λ values, the described backpropagation update drives both λ values toward zero, making Ltotal collapse to zero and switching off the physics residual. The claimed 'dynamic balance' mechanism is thus self-defeating: the adaptive weights, by their own update rule, eliminate the physics term they are meant to weight. The Table 6 ablation gap (0.2515 vs 0.2880 mm) then cannot be attributed to the SPI loss as written.

full rationale

The central forward derivation is not circular: the final prediction is ppred = ptheory + α·Δp (Eq. 9), where ptheory is the DH-model output and Δp is learned from measured ground-truth coordinates through Ldata (Eq. 10). The sparse attention mask is a newly designed matrix (Eq. 4) inspired by, but not identical to, the externally cited Body Transformer [36]. The SPI distance-matrix term (Eq. 16) is a regularizer using the same kinematic prior as the forward branch; that is a modeling choice rather than a reduction of the prediction to its inputs. The DNN and GPSO-DNN comparisons are external benchmarks on the same test split, so the headline 35.16% improvement has independent content. The by-construction elements are secondary: the 0.01 mm inverse-convergence figure equals the early-stopping threshold, and the 97.3% spatial-similarity figure restates the SPI loss. Separately, the Eq. (18) adaptive lambda update is a self-defeating mechanism because its gradients are the nonnegative losses, driving λ→0; this is a correctness risk in the ablation attribution, not a circularity of the forward model. No load-bearing self-citation chain or imported uniqueness theorem is present. Overall score 4 reflects partial by-construction reporting while the main forward result remains independent.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The paper's central claims depend on the nominal DH model, the measurement chain, the stability of the adaptive loss weighting, the smoothness of the learned forward model, and the representativeness of the single data split. No new physical entities are introduced.

free parameters (2)
  • alpha (fusion scale) = trainable, initialized to 0.1, final value not reported
    Controls the blend of theoretical D-H prediction and learned residual in Eq. 9; fitted to data.
  • lambda_data and lambda_physics (loss weights) = trainable, final values not reported
    Adaptive weights in Eq. 18; as described they are updated by gradient descent and are prone to collapsing to zero.
assumptions (5)
  • domain assumption The nominal DH parameters in Table 1 correctly describe the UR5's ideal forward kinematics, so p_theory is a valid reference for both residual learning and the distance-matrix loss.
    The physics branch (Section 2.1.1) uses these fixed values; no calibration of the DH parameters is performed, so any bias in these values propagates into the theoretical positions and the SPI loss reference.
  • domain assumption The TrackScan Sharp 3D scanning system and the SVD-based base-to-world calibration (Section 3.1) provide ground-truth positions accurate enough to support errors at the 0.1 mm scale.
    Scanner specifications are cited, but calibration residual errors are not reported; the entire error dataset depends on this measurement chain.
  • ad hoc to paper The Euclidean distance-matrix loss (Eq. 16) improves prediction quality beyond ordinary data fitting.
    The ablation (Table 6) shows a 12% improvement, but the mechanism is empirical; no proof or analysis explains why aligning normalized pairwise distances with the theory generalizes.
  • domain assumption The trained BoTER model is smooth enough that Adam-based gradient descent on the input angles reliably finds correct joint compensations.
    Section 2.3 optimizes joint angles through the frozen network; no analysis of local minima or convergence guarantees is provided.
  • ad hoc to paper A single 8:1:1 split with seed 7 and a single training run with seed 139 are representative of the model's generalization.
    No multiple seeds or cross-validation are reported, so the headline 35.16% improvement could be split-specific.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SPI-BoTER: Error Compensation for Industrial Robots via Sparse Attention Masking and Hybrid Loss with Spatial-Physical Information." pith.science (2026). https://pith.science/paper/JRSX2OER

@misc{pith2026250622788,
  author       = {Pith},
  title        = {Pith review of: SPI-BoTER: Error Compensation for Industrial Robots via Sparse Attention Masking and Hybrid Loss with Spatial-Physical Information},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JRSX2OER}},
  note         = {Machine review of arXiv:2506.22788}
}
read the original abstract

The widespread application of industrial robots in fields such as cutting and welding has imposed increasingly stringent requirements on the trajectory accuracy of end-effectors. However, current error compensation methods face several critical challenges, including overly simplified mechanism modeling, a lack of physical consistency in data-driven approaches, and substantial data requirements. These issues make it difficult to achieve both high accuracy and strong generalization simultaneously. To address these challenges, this paper proposes a Spatial-Physical Informed Attention Residual Network (SPI-BoTER). This method integrates the kinematic equations of the robotic manipulator with a Transformer architecture enhanced by sparse self-attention masks. A parameter-adaptive hybrid loss function incorporating spatial and physical information is employed to iteratively optimize the network during training, enabling high-precision error compensation under small-sample conditions. Additionally, inverse joint angle compensation is performed using a gradient descent-based optimization method. Experimental results on a small-sample dataset from a UR5 robotic arm (724 samples, with a train:test:validation split of 8:1:1) demonstrate the superior performance of the proposed method. It achieves a 3D absolute positioning error of 0.2515 mm with a standard deviation of 0.15 mm, representing a 35.16\% reduction in error compared to conventional deep neural network (DNN) methods. Furthermore, the inverse angle compensation algorithm converges to an accuracy of 0.01 mm within an average of 147 iterations. This study presents a solution that combines physical interpretability with data adaptability for high-precision control of industrial robots, offering promising potential for the reliable execution of precision tasks in intelligent manufacturing.

Figures

Figures reproduced from arXiv: 2506.22788 by the authors.

Figure 1
Figure 1. Schematic diagram of the complete error compensation process for industrial robots in this study [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Schematic Diagram of the BoTER Model Architecture [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Kinematics properties for UR robot We allow each joint to attend to its directly connected parent/child joints—i.e., neighboring joints are mutually visible. However, in contrast to the masking rules in Body Transformer [36], we design a hierarchical attention 7 [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (14 more)
Figure 4
Figure 4. Figure 4: Simulation of one batch of training samples optimized using the SPI loss function. [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Comparison of distance matrices for a batch of test samples. [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 6
Figure 6. Figure 6: Inverse Angle Solving Algorithm Pipeline [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]
Figure 7
Figure 7. Figure 7: Experiments of error data collection system achieves a high-precision measurement volume of up to 135 cubic meters, with a maximum scanning range extending to 8.5 meters. Detailed specifications are provided in [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]
Figure 8
Figure 8. Figure 8: 3D sample distribution (1) Initialization: The BoTER model weights were initialized using a random seed of 139. Based on the kinematic structure of the UR5 robotic arm, a sparse self-attention mask was designed accordingly. (2) Loss Function Definition: An SPI loss fun…
Figure 9
Figure 9. Figure 9: Training process of SPI-BoTER • Input/Output Dimensions: 6-dimensional joint angle input and 3-dimensional end-effector position output. • Transformer Encoder: Input dimension dmodel = 126, number of encoder layers Nlayer = 4, number of attention heads nhead = 9. 15 […
Figure 10
Figure 10. Figure 10: True vs. Predicted Error Scatter Chart on the X-axis (Test Set) [PITH_FULL_IMAGE:figures/full_fig_p017_10.png]
Figure 11
Figure 11. Figure 11: True vs. Predicted Error Scatter Chart on the Y-axis (Test Set) [PITH_FULL_IMAGE:figures/full_fig_p018_11.png]
Figure 12
Figure 12. Figure 12: True vs. Predicted Error Scatter Chart on the Z-axis (Test Set) [PITH_FULL_IMAGE:figures/full_fig_p018_12.png]
Figure 13
Figure 13. Figure 13: Error Distribution on the Test Set Figures 10–12 illustrate the scatter plots of predicted versus actual errors for the X, Y and Z axes of the test set. These plots compare the differences between the predicted and theoretical coordinates with those between the measur…
Figure 14
Figure 14. Figure 14: Performance Comparison between DNN and SPI-BoTER across MAE, MSE, RMSE, and [PITH_FULL_IMAGE:figures/full_fig_p019_14.png]
Figure 15
Figure 15. Figure 15: Performance Comparison of Ablation Groups for Each Loss Term of SPI [PITH_FULL_IMAGE:figures/full_fig_p021_15.png]
Figure 16
Figure 16. Figure 16: Performance Comparison of SPI-BoTER Model Structure Ablation Groups [PITH_FULL_IMAGE:figures/full_fig_p022_16.png]
Figure 17
Figure 17. Figure 17: Position Error of 50 Randomly Sampled Test Points [PITH_FULL_IMAGE:figures/full_fig_p023_17.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

46 extracted references · 27 canonical work pages

  1. [1]

    In Situ Reconfiguration of Assembling Pattern for Modular Continuum Robots

    Zhang J, Cai J, Ma K, et al. In Situ Reconfiguration of Assembling Pattern for Modular Continuum Robots. Journal of Bionic Engineering, 2024, 21(4): 1695–1706. https://doi.org/10.1007/s42235-024-00523-5

    work page doi:10.1007/s42235-024-00523-5 2024

  2. [2]

    A welding task data model for intelligent process planning of robotic welding

    Shen W, Hu T, Zhang C, et al. A welding task data model for intelligent process planning of robotic welding. Robotics and Computer- Integrated Manufacturing, 2020, 64: 101934. https://doi.org/10.1016/j.rcim.2020.101934

    arXiv 2020

  3. [3]

    Chatter stability in robotic milling

    Cordes M, Hintze W, Altintas Y . Chatter stability in robotic milling. Robotics and Computer-Integrated Manufacturing, 2019, 55: 11–18. https://doi.org/10.1016/j.rcim.2018.07.004

    work page doi:10.1016/j.rcim.2018.07.004 2019

  4. [4]

    Pose accuracy improvement in robotic machining by visually-guided method and experimental investigation

    Li B, Li Y , Tian W, et al. Pose accuracy improvement in robotic machining by visually-guided method and experimental investigation. Robotics and Autonomous Systems, 2023, 164: 104416. https://doi.org/10.1016/j.robot.2023.104416

    work page arXiv 2023

  5. [5]

    Positional error compensation for aviation drilling robot based on Bayesian linear regression

    Chen D, Lv P, Xue L, et al. Positional error compensation for aviation drilling robot based on Bayesian linear regression. Engineering Applications of Artificial Intelligence, 2024, 127: 107263. https://doi.org/10.1016/j.engappai.2023.107263

    arXiv 2024

  6. [6]

    Error compensation technology and its application progress of an industrial robot, 2022, 43(5): 627142–627142

    Liao Wenhe TW, Li Bo, Li Pengcheng, Zhang Wei, Li Yufei. Error compensation technology and its application progress of an industrial robot, 2022, 43(5): 627142–627142. https://doi.org/10.7527/s1000-6893.2022.27142

    work page arXiv 2022

  7. [7]

    Study of Orientation Error on Robot End E ffector and V olumetric Error of Articulated Robot, 2019, 9(23): 5149

    OH YT. Study of Orientation Error on Robot End E ffector and V olumetric Error of Articulated Robot, 2019, 9(23): 5149. https://doi. org/10.3390/app9235149 24

    work page doi:10.3390/app9235149 2019

  8. [8]

    A Practical Method to Improve Absolute Positioning Accuracy of Industrial Robot

    Shi G, Zhao S, Hu B. A Practical Method to Improve Absolute Positioning Accuracy of Industrial Robot. Journal of Physics: Conference Series, 2020, 1453(1): 012121. https://doi.org/10.1088/1742-6596/1453/1/012121

    work page doi:10.1088/1742-6596/1453/1/012121 2020

Show all 46 references
  1. [9]

    A Data-Driven Approach for Online Path Correction of Industrial Robots Using Modified Flexible Dynamics Model and Disturbance State Observer

    Lin Y , Zhao H, Ding H. A Data-Driven Approach for Online Path Correction of Industrial Robots Using Modified Flexible Dynamics Model and Disturbance State Observer. IEEE /ASME Transactions on Mechatronics, 2023, 28(6): 3410–3421. https://doi.org/10. 1109/TMECH.2023.3267980

    2023

  2. [10]

    Hybrid o ffline programming method for robotic welding systems

    Zheng C, An Y , Wang Z, et al. Hybrid o ffline programming method for robotic welding systems. Robotics and Computer-Integrated Manu- facturing, 2022, 73: 102238. https://doi.org/10.1016/j.rcim.2021.102238

    2022

  3. [11]

    IK-Geo: Unified robot inverse kinematics using subproblem decomposition

    Elias AJ, Wen JT. IK-Geo: Unified robot inverse kinematics using subproblem decomposition. Mechanism and Machine Theory, 2025, 209: 105971. https://doi.org/10.1016/j.mechmachtheory.2025.105971

    2025

  4. [12]

    A Kinematic Notation for Lower-Pair Mechanisms

    Denavit J, Hartenberg RS. A Kinematic Notation for Lower-Pair Mechanisms. Journal of Applied Mechanics, 1955, 22: 215–221. https: //doi.org/10.1115/1.4011045

    1955 doi

  5. [13]

    Evaluation of Kinematic and Compliance Calibration of Serial Articulated Industrial Manipulators

    Ibaraki S, Theissen NA, Archenti A, et al. Evaluation of Kinematic and Compliance Calibration of Serial Articulated Industrial Manipulators. International Journal of Automation Technology, 2021, 15(5): 567–580. https://doi.org/10.20965/ijat.2021.p0567

    2021 doi

  6. [14]

    A calibration method for spatial error of industrial robot under variable load conditions based on double-ball rotary structure and three contact displacement sensors

    Yang P, Luo K, Wu R. A calibration method for spatial error of industrial robot under variable load conditions based on double-ball rotary structure and three contact displacement sensors. Precision Engineering, 2024, 85: 291–303. https://doi.org/10.1016/j. precisioneng.2023.10.016

    2024 doi

  7. [15]

    Kinematic Modeling and Analysis of Robotic Axial Tilt Errors

    Feng X, Zhang X, Tu D. Kinematic Modeling and Analysis of Robotic Axial Tilt Errors. In: Lan X, Mei X, Jiang C, Zhao F, Tian Z (Eds.) Intelligent Robotics and Applications. Springer Nature Singapore, Singapore, 2025: 154–165. https://doi.org/10.1007/ 978-981-96-0771-6_12

    2025

  8. [16]

    Modelling Posture-Dependent Denavit-Hartenberg Errors of A Six-Axis Robotic Manipulator

    Cui TH, Ibaraki S. Modelling Posture-Dependent Denavit-Hartenberg Errors of A Six-Axis Robotic Manipulator. In: Proceedings of 2024 International Symposium on Flexible Automation (ISFA 2024), 2024. https://doi.org/10.1115/ISFA2024-140901

    2024 doi

  9. [17]

    A Two-Stage Parameter Identification Method and Compensation Verification for Heavy Load Robot

    Wang Z, Mao C, Chen Z, et al. A Two-Stage Parameter Identification Method and Compensation Verification for Heavy Load Robot. In: 2019 4th International Conference on Control, Robotics and Cybernetics (CRC), 2019: 38–41. https://doi.org/10.1109/CRC.2019.00017

    2019

  10. [18]

    Nonlinear Error Compensation for Load Cells Based on the Optimal Neural Network With an Augmented Lagrange Multiplier

    Lin H, Wang L, Yu J, et al. Nonlinear Error Compensation for Load Cells Based on the Optimal Neural Network With an Augmented Lagrange Multiplier. IEEE Transactions on Instrumentation and Measurement, 2015, 64(11): 2850–2862. https://doi.org/10.1109/TIM.2015. 2433651

    2015 doi

  11. [19]

    Deep learning-based interpretable prediction and compensation method for improving pose accuracy of parallel robots

    Zhu X, Zhang H, Liu Z, et al. Deep learning-based interpretable prediction and compensation method for improving pose accuracy of parallel robots. Expert Systems with Applications, 2025, 268: 126289. https://doi.org/10.1016/j.eswa.2024.126289

    2025

  12. [20]

    Error Prediction for Large Optical Mirror Processing Robot Based on Deep Learning

    Jin Z, Cheng G, Xu S, et al. Error Prediction for Large Optical Mirror Processing Robot Based on Deep Learning. Journal of Mechanical Engineering, 2022, 68(3): 175. https://doi.org/10.5545/sv-jme.2021.7455

    2022

  13. [21]

    Improving Industrial Robot Positioning Accuracy to the Microscale Using Machine Learning Method, 2022, 10(10): 940

    Bucinskas V , Dzedzickis A, Sumanas M, et al. Improving Industrial Robot Positioning Accuracy to the Microscale Using Machine Learning Method, 2022, 10(10): 940. https://doi.org/10.3390/machines10100940

    2022 doi

  14. [22]

    Predicting Positioning Error and Finding Features for Large Industrial Robots Based on Deep Learning

    Kato D, Yoshitsugu K, Hirogaki T, et al. Predicting Positioning Error and Finding Features for Large Industrial Robots Based on Deep Learning. International Journal of Automation Technology, 2021, 15(2): 206–214. https://doi.org/10.20965/ijat.2021.p0206

    2021 doi

  15. [23]

    High-accuracy prediction and compensation of industrial robot stiffness deformation

    Ye C, Yang J, Ding H. High-accuracy prediction and compensation of industrial robot stiffness deformation. International Journal of Mechan- ical Sciences, 2022, 233: 107638. https://doi.org/10.1016/j.ijmecsci.2022.107638

    2022

  16. [24]

    Error compensation of industrial robot based on deep belief network and error similarity

    Wang W, Tian W, Liao W, et al. Error compensation of industrial robot based on deep belief network and error similarity. Robotics and Computer-Integrated Manufacturing, 2022, 73: 102220. https://doi.org/10.1016/j.rcim.2021.102220

    2022

  17. [25]

    Introduction of a Framework for the Integration of a Kinematic Robot Arm Model in an Artificial Neural Net- work - Extended Kalman Filter Approach

    Horvath S, Neuner H. Introduction of a Framework for the Integration of a Kinematic Robot Arm Model in an Artificial Neural Net- work - Extended Kalman Filter Approach. Journal of Intelligent & Robotic Systems, 2024, 110(4). https://doi.org/10.1007/ s10846-024-02164-6

    2024

  18. [26]

    A degradation assessment method for industrial robot in flexible manufacturing systems based on multiple physics-informed neural network

    Wang S, Tao J, Jiang Q, et al. A degradation assessment method for industrial robot in flexible manufacturing systems based on multiple physics-informed neural network. Mechanical Systems and Signal Processing, 2025, 233: 112793.https://doi.org/10.1016/j.ymssp. 2025.112793

    2025

  19. [27]

    Roughness Prediction of Jointed Plain Concrete Pavement Using Physics Informed Neural Networks

    Pasupunuri SK, Thom N, Li L. Roughness Prediction of Jointed Plain Concrete Pavement Using Physics Informed Neural Networks. Trans- portation Research Record: Journal of the Transportation Research Board, 2024, 2678(11): 1733–1746. https://doi.org/10.1177/ 03611981241245991

    2024

  20. [28]

    Physics informed neural network for parameter identification and boundary force estimation of compliant and biome- chanical systems

    Li W, Lee K-M. Physics informed neural network for parameter identification and boundary force estimation of compliant and biome- chanical systems. International Journal of Intelligent Robotics and Applications, 2021, 5(3): 313–325. https://doi.org/10.1007/ s41315-021-00196-x

    2021

  21. [29]

    A PINN-Based Friction-Inclusive Dynamics Modeling Method for Industrial Robots

    Hu H, Shen Z, Zhuang C. A PINN-Based Friction-Inclusive Dynamics Modeling Method for Industrial Robots. IEEE Transactions on Indus- trial Electronics, 2025, 72(5): 5136–5144. https://doi.org/10.1109/tie.2024.3476977

    2025

  22. [30]

    Physics-Informed Neural Network for Model Prediction and Dynamics Parameter Identification of Collaborative Robot Joints

    Yang X, Du Y , Li L, et al. Physics-Informed Neural Network for Model Prediction and Dynamics Parameter Identification of Collaborative Robot Joints. IEEE Robotics and Automation Letters, 2023, 8(12): 8462–8469. https://doi.org/10.1109/lra.2023.332960

    2023

  23. [31]

    Attention is All you Need

    Vaswani A, Shazeer NM, Parmar N, et al. Attention is All you Need. arXiv:1706.03762 cs.ML, 2017. https://doi.org/10.48550/ arXiv.1706.03762

    2017 doi arXiv

  24. [32]

    Learning to Improve Operational E fficiency from Pose Error Estimation in Robotic Pollination

    Chen J, Xiao J, Yang M, et al. Learning to Improve Operational E fficiency from Pose Error Estimation in Robotic Pollination. Electronics, 2024, 13(15): 3070. https://doi.org/10.3390/electronics13153070

    2024 doi

  25. [33]

    DDETR-SLAM: A Transformer-Based Approach to Pose Optimization in Dynamic Environments

    Li F, Liu Y , Wang G, et al. DDETR-SLAM: A Transformer-Based Approach to Pose Optimization in Dynamic Environments. arXiv:2306.08658 cs.RO, 2023. https://doi.org/10.21203/rs.3.rs-2965479/v1

    2023 arXiv

  26. [34]

    A CNN-LSTM framework for flight delay prediction

    Li Q, Guan X, Liu J. A CNN-LSTM framework for flight delay prediction. Expert Systems with Applications, 2023, 227: 120287. https: //doi.org/10.1016/j.eswa.2023.120287

    2023

  27. [35]

    Semantically-Enhanced Feature Extraction with CLIP and Transformer Networks for Driver Fatigue Detection, 2024, 24(24): 7948

    Gao Z, Chen X, Xu J, et al. Semantically-Enhanced Feature Extraction with CLIP and Transformer Networks for Driver Fatigue Detection, 2024, 24(24): 7948. https://doi.org/10.3390/s24247948

    2024 doi

  28. [36]

    Body transformer: Leveraging robot embodiment for policy learning

    Sferrazza C, Huang D-M, Liu F, et al. Body transformer: Leveraging robot embodiment for policy learning. arXiv:2408.06316 cs.RO, 2024. 25 https://doi.org/10.48550/arXiv.2408.06316

    2024 doi arXiv

  29. [37]

    iTransformer: Inverted Transformers Are Effective for Time Series Forecasting

    Liu Y , Hu T, Zhang H, et al. iTransformer: Inverted Transformers Are Effective for Time Series Forecasting. arXiv:2310.06625 cs.LG, 2023. https://doi.org/10.48550/arXiv.2310.06625

    2023 doi arXiv

  30. [38]

    Multi-objective optimization of high-power fiber laser cutting process using data augmentation-based ANN-Adam modelJ

    Liu Y , Yin Y , Zhang S. Multi-objective optimization of high-power fiber laser cutting process using data augmentation-based ANN-Adam modelJ. Optical Fiber Technology, 2024, 87: 103875. ---

    2024

  31. [39]

    Research on Kinematics Analysis and Trajectory Planning of UR5 RobotC //Journal of Physics: Conference Series

    Wen Z, Liu F, Dou X, et al. Research on Kinematics Analysis and Trajectory Planning of UR5 RobotC //Journal of Physics: Conference Series. IOP Publishing, 2022, 2396(1): 012046. https://doi.org/10.1088/1742-6596/2396/1/012046

    2022 doi

  32. [40]

    RELU-function and derived function reviewC//SHS Web of Conferences

    Bai Y . RELU-function and derived function reviewC//SHS Web of Conferences. EDP Sciences, 2022, 144: 02006. https://doi.org/10. 1051/shsconf/202214402006

    2022

  33. [41]

    Euclidean distance matrices: essential theory, algorithms, and applicationsJ

    Dokmanic I, Parhizkar R, Ranieri J, et al. Euclidean distance matrices: essential theory, algorithms, and applicationsJ. IEEE Signal Processing Magazine, 2015, 32(6): 12–30. https://doi.org/10.1109/MSP.2015.2443937

    2015

  34. [42]

    The coe fficient of determination R-squared is more informative than SMAPE, MAE, MAPE, MSE and RMSE in regression analysis evaluationJ

    Chicco D, Warrens MJ, Jurman G. The coe fficient of determination R-squared is more informative than SMAPE, MAE, MAPE, MSE and RMSE in regression analysis evaluationJ. PeerJ Computer Science, 2021, 7: e623. https://doi.org/10.7717/peerj-cs.623

    2021 doi

  35. [43]

    Positioning error compensation of an industrial robot using neural networks and experimental study

    Li B, Tian W, Zhang C, Hua F, Cui G, Li Y . Positioning error compensation of an industrial robot using neural networks and experimental study. Chinese Journal of Aeronautics, 2022, 35(2): 346–360. https://doi.org/10.1016/j.cja.2021.03.027

    2022 doi

  36. [44]

    Native sparse attention: Hardware-aligned and natively trainable sparse attentionJ

    Yuan J, Gao H, Dai D, et al. Native sparse attention: Hardware-aligned and natively trainable sparse attentionJ. arXiv preprint arXiv:2502.11089, 2025. https://doi.org/10.48550/arXiv.2502.11089

    2025 doi arXiv

  37. [45]

    Generating long sequences with sparse transformersJ

    Child R, Gray S, Radford A, et al. Generating long sequences with sparse transformersJ. arXiv preprint arXiv:1904.10509, 2019. https: //doi.org/10.48550/arXiv.1904.10509

    1904 doi arXiv

  38. [46]

    Robot positioning error compensation method based on deep neural networkC //Journal of Physics: Conference Series

    Hu J, Hua F, Tian W. Robot positioning error compensation method based on deep neural networkC //Journal of Physics: Conference Series. IOP Publishing, 2020, 1487(1): 012045. https://doi.org/10.1088/1742-6596/1487/1/012045 26

    2020 doi

Pith tools

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