Hybrid Dynamics Modeling of Industrial Robots

I propose a hybrid dynamics modeling framework for industrial robot manipulators that combines a physically interpretable baseline model with a spatiotemporal graph neural network residual compensator. The central motivation is that a purely physics-based inverse dynamics model of an industrial robot cannot fully describe friction nonlinearity, joint compliance, transmission hysteresis, actuator delay, and inter-joint coupling. Therefore, I first identify a physically feasible baseline model from measured trajectory data, and then I learn the remaining structured residual with a spatiotemporal graph neural network equipped with a Kolmogorov-Arnold Network regression head. I call this compensation model STG-KAN. Across four trajectory-level repeated experiments, the proposed hybrid model achieves an overall root-mean-square error of \(15.25 \pm 1.28\,\mathrm{N\cdot m}\), which is \(33.75\%\) lower than the baseline physical model and lower than several deep-learning compensation baselines. The results indicate that an industrial robot can retain physical interpretability while gaining substantial data-driven accuracy when spatial topology and temporal hysteresis are modeled jointly.

1. Introduction

Industrial robot applications increasingly require accurate trajectory tracking, reliable force control, safe interaction, and fast state estimation. In all of these tasks, the inverse dynamics model of the industrial robot is a fundamental component. If the model predicts joint torque poorly, then feedforward control, collision detection, and model-based monitoring all degrade. The difficulty is that the true dynamics of an industrial robot include not only rigid-body inertial, Coriolis, centrifugal, and gravitational effects but also friction, flexibility, backlash, actuator dynamics, and direction-dependent coupling. A rigid-body model with identified parameters is therefore accurate in trend but not in every local segment. This gap is especially visible near velocity reversals, torque peaks, and fast transients.

Parameter identification has been the classical route for improving the physical model of an industrial robot. The rigid-body dynamics can be written in a linear parameterized form, and the parameters can be estimated from excitation trajectories. However, standard parameters are often linearly dependent, so base parameters must be extracted. Moreover, the identified model may violate physical feasibility if the inertia matrix is not positive definite or if mass parameters are negative. I therefore formulate identification as a constrained convex problem with physical feasibility inequalities. This gives a baseline model that is stable, interpretable, and suitable for residual learning.

Data-driven compensation is another route. Neural networks can approximate unmodeled dynamics, but a generic multilayer perceptron or recurrent network does not explicitly represent the serial topology of an industrial robot. The dynamics of an industrial robot propagate from the base toward the end effector in the kinematics chain, while inertial and load effects accumulate backward from the end effector toward the base. This directional topology is not captured by a plain temporal model. In addition, friction and hysteresis introduce temporal dependencies that span multiple sampling instants. I therefore design a spatiotemporal graph neural network that represents directional coupling through multi-relation graphs and represents temporal hysteresis through multiscale causal dilated convolution and temporal attention. I further replace the usual multilayer regression head with a KAN head to improve nonlinear residual fitting.

The contributions of my work are as follows. First, I build a hybrid inverse dynamics framework for an industrial robot in which a physically feasible baseline and a data-driven residual compensator are trained in a coordinated manner. Second, I construct a multi-relation graph that includes downward propagation, upward feedback, nonlocal coupling, and a learnable prior correction, so that the industrial robot topology is explicitly modeled. Third, I combine multiscale causal temporal convolution, gated fusion, and temporal attention to capture friction reversal and actuator hysteresis. Fourth, I introduce a KAN regression head for residual prediction. Fifth, I validate the model on a six-degree-of-freedom industrial robot platform with repeated trajectory-level experiments and ablation studies.

2. Physical Dynamics of an Industrial Robot

For a serial industrial robot with \(n\) joints, the rigid-body inverse dynamics can be expressed by the Newton-Euler formulation as

$$
\tau = M(q)\ddot q + C(q,\dot q)\dot q + G(q) + \tau_f,
$$

where \(q,\dot q,\ddot q \in \mathbb{R}^n\) are joint position, velocity, and acceleration, respectively; \(M(q)\in\mathbb{R}^{n\times n}\) is the inertia matrix; \(C(q,\dot q)\in\mathbb{R}^{n\times n}\) is the Coriolis and centrifugal matrix; \(G(q)\in\mathbb{R}^n\) is the gravity vector; \(\tau\in\mathbb{R}^n\) is the joint drive torque; and \(\tau_f\in\mathbb{R}^n\) is the friction torque. For an industrial robot, the friction term is often represented by a Coulomb-viscous model:

$$
\tau_f = F_c \operatorname{sgn}(\dot q) + F_v \dot q,
$$

where \(F_c\) and \(F_v\) are Coulomb and viscous friction coefficients. Although this model captures the dominant friction trend, it does not represent Stribeck effects, presliding displacement, direction-dependent friction, or transmission hysteresis. These unmodeled effects become residual targets for the data-driven compensator.

Because the rigid-body dynamics of an industrial robot are linear in the inertial parameters, I can write

$$
\tau = Y(q,\dot q,\ddot q)P,
$$

where \(Y(q,\dot q,\ddot q)\in\mathbb{R}^{n\times s}\) is the regressor matrix and \(P\in\mathbb{R}^{s}\) is the standard parameter vector. Since the standard parameters contain structural coupling and linear dependence, the regressor often has rank deficiency. I therefore use inertial parameter regrouping to obtain a full-column-rank regressor \(Y_b\) and a base parameter vector \(P_b\):

$$
\tau = Y_b(q,\dot q,\ddot q)P_b.
$$

This base-parameter form is the foundation of the baseline physical model for the industrial robot.

3. Excitation Trajectory and Physical Feasibility

To identify the baseline dynamics of an industrial robot, the excitation trajectory must sufficiently excite the inertial, centrifugal, Coriolis, gravity, and friction effects while respecting joint limits. I use a finite Fourier series for each joint:

$$
q_i(t)=q_{i0}+\sum_{\ell=1}^{N}\frac{a_{\ell,i}}{\ell\omega_f}\sin(\ell\omega_f t)-\frac{b_{\ell,i}}{\ell\omega_f}\cos(\ell\omega_f t),
$$

where \(N\) is the Fourier order, \(\omega_f\) is the base frequency, \(q_{i0}\) is the initial offset, and \(a_{\ell,i}\) and \(b_{\ell,i}\) are Fourier coefficients. I optimize the trajectory by minimizing the condition number of the regressor matrix:

$$
\min_{a,b}\operatorname{cond}(Y),
$$

subject to position, velocity, acceleration, and smooth start-stop constraints. The optimization is summarized in Table 1.

Constraint Mathematical form Purpose
Joint position limit \(|q_i(t)|\le q_{i,\max}\) Avoid mechanical collision
Joint velocity limit \(|\dot q_i(t)|\le v_{i,\max}\) Respect actuator and safety limits
Joint acceleration limit \(|\ddot q_i(t)|\le a_{i,\max}\) Limit inertial excitation
Start-stop smoothness \(q_i(t_0)=q_i(t_f)\), \(\dot q_i(t_0)=\dot q_i(t_f)=0\), \(\ddot q_i(t_0)=\ddot q_i(t_f)=0\) Enable periodic and repeatable identification
Regressor conditioning \(\min \operatorname{cond}(Y)\) Improve numerical identifiability

For physical consistency, I impose positive mass and positive-definite inertial conditions on each link of the industrial robot. For link \(k\), the mass satisfies \(m_k>0\), and the pseudo-inertia condition can be written as a linear matrix inequality:

$$
D_k(\theta)=
\begin{bmatrix}
I_k & S(p_k)\\
S(p_k)^T & m_k I
\end{bmatrix}
\succ 0,
$$

where \(I_k\) is the link inertia tensor and \(S(p_k)\) is the skew-symmetric operator associated with the first moment \(p_k\). The constrained identification problem is then

$$
\min_{\theta}\|Y_b\theta-\tau_{\mathrm{vec}}\|_2^2
\quad
\text{s.t.}\quad
D_k(\theta)\succ 0,\quad f_k\ge 0,
$$

where \(\tau_{\mathrm{vec}}\) is the stacked measured torque vector and \(f_k\) denotes friction parameters. I solve this problem as a semidefinite program. The resulting baseline model is physically feasible and provides a stable prediction for the industrial robot.

4. Residual Learning Problem

After the baseline physical model is obtained, I define the residual at time \(t\) as

$$
r(t)=\tau_{\mathrm{actual}}(t)-\tau_{\mathrm{pred}}(t),
$$

where \(\tau_{\mathrm{actual}}\) is the measured joint torque and \(\tau_{\mathrm{pred}}\) is the baseline prediction. For each joint \(i\), I construct a node feature vector:

$$
z_i(t)=
\begin{bmatrix}
q_i(t), & \dot q_i(t), & \ddot q_i(t), & \operatorname{sgn}(\dot q_i(t))
\end{bmatrix}^T.
$$

The sign term informs the network about friction direction, which is important near velocity reversals. For an industrial robot with \(N\) joints, the system-level feature at time \(t\) is

$$
Z(t)=
\begin{bmatrix}
z_1(t)^T, & z_2(t)^T, & \cdots, & z_N(t)^T
\end{bmatrix}^T.
$$

I use a sliding time window of length \(W=20\) and step 1 to form the input tensor:

$$
X_t=
\begin{bmatrix}
Z(t-W+1), & Z(t-W+2), & \cdots, & Z(t)
\end{bmatrix}.
$$

The input therefore contains both instantaneous joint states and historical context. Table 2 summarizes the residual learning input structure.

Element Definition Role in industrial robot compensation
Joint position \(q_i(t)\) Captures configuration-dependent gravity and inertia
Joint velocity \(\dot q_i(t)\) Captures viscous friction and Coriolis effects
Joint acceleration \(\ddot q_i(t)\) Captures inertial torque and fast transients
Velocity sign \(\operatorname{sgn}(\dot q_i(t))\) Captures Coulomb friction direction
Sliding window \(W=20\) Captures hysteresis and temporal dependency

5. STG-KAN Hybrid Architecture

The proposed STG-KAN model compensates the baseline residual of the industrial robot by jointly modeling spatial topology and temporal dynamics. The architecture consists of four parts: an input projection, a multi-relation graph convolution module, a multiscale causal temporal encoding module, and a KAN residual regression head. I describe each part below.

5.1 Multi-Relation Graph Model of the Industrial Robot

The serial structure of an industrial robot induces directional dynamic propagation. Motion states propagate from the base to the end effector, while inertial and load effects propagate backward. I therefore define three physical relation graphs: a downward propagation graph \(A_{\mathrm{down}}\), an upward feedback graph \(A_{\mathrm{up}}\), and a coupling graph \(A_{\mathrm{cpl}}\). For a six-joint industrial robot, \(A_{\mathrm{down}}\) follows the chain \(1\to2\to\cdots\to6\), while \(A_{\mathrm{up}}\) follows the reverse direction. Both are row-normalized. The coupling graph connects all joint pairs and represents nonlocal dynamic coupling.

Because non-adjacent links also contribute to torque through inertial coupling and load transmission, I introduce a distance-decay prior graph:

$$
A^{\mathrm{prior}}_{ij}=
\begin{cases}
1, & i=j,\\
\exp\left(-\dfrac{|i-j|}{\sigma}\right), & i\ne j,
\end{cases}
$$

where \(\sigma>0\) is a decay coefficient. I use \(\sigma=1.2\). To allow the model to learn cross-joint relations that the prior cannot express, I add a learnable correction:

$$
\Delta A = s\tanh\left(\frac{BB^T}{2}\right),
$$

where \(B\) is a learnable parameter matrix and \(s=0.3\) is a scaling factor. The final coupling adjacency is

$$
\hat A_{\mathrm{cpl}}=
D^{-1/2}\left(A^{\mathrm{prior}}+\Delta A\right)D^{-1/2},
$$

where \(D\) is the degree matrix. The multi-relation graph design is summarized in Table 3.

Graph Symbol Physical meaning Direction
Downward propagation \(A_{\mathrm{down}}\) Base-to-tip kinematic and dynamic propagation \(1\to2\to\cdots\to N\)
Upward feedback \(A_{\mathrm{up}}\) Tip-to-base inertial and load accumulation \(N\to\cdots\to2\to1\)
Coupling graph \(A_{\mathrm{cpl}}\) Nonlocal dynamic interaction among joints All joint pairs
Prior graph \(A^{\mathrm{prior}}\) Distance-decay coupling prior Symmetric
Learned correction \(\Delta A\) Adaptive residual relation Symmetric

For an input feature tensor \(H\in\mathbb{R}^{J\times T\times C}\), the multi-relation graph convolution is

$$
G(H)=
\phi_s(H)
+\alpha_d\phi_d(\hat A_{\mathrm{down}}H)
+\alpha_u\phi_u(\hat A_{\mathrm{up}}H)
+\alpha_c\phi_c(\hat A_{\mathrm{cpl}}H),
$$

where \(\phi_s,\phi_d,\phi_u,\phi_c\) are linear mappings implemented by convolution, and the gates are

$$
\alpha_d=\sigma(g_d),\quad
\alpha_u=\sigma(g_u),\quad
\alpha_c=\sigma(g_c),
$$

with learnable scalars \(g_d,g_u,g_c\) and sigmoid function \(\sigma(\cdot)\). This gating mechanism allows the industrial robot model to adaptively select the contribution of each relation during different motion phases.

5.2 Multiscale Causal Temporal Encoding

The residual of an industrial robot is temporally correlated because friction, compliance, and actuator hysteresis depend on past states. I use causal dilated convolution to encode historical dependence while preventing future information leakage. The basic form is

$$
T_d(h)(t)=\sum_{m=0}^{K-1}\Theta_m h(t-md),
$$

where \(K\) is the kernel width, \(d\) is the dilation rate, and \(\Theta_m\) are convolution parameters. I use two parallel branches with kernel widths 3 and 5. Three spatiotemporal blocks use dilation rates \(1,2,4\), and each branch has 64 output channels. Left-side causal padding keeps the sequence length unchanged.

To capture both short-term fluctuations caused by friction reversal and long-term dependency caused by drive hysteresis, I fuse multiple temporal branches:

$$
z=\sum_{r=1}^{R}\beta_r z_r,
$$

$$
\beta_r=
\frac{\exp(\gamma_r)}
{\sum_{k=1}^{R}\exp(\gamma_k)},
$$

where \(\gamma_r\) are learnable branch parameters. I also apply a gated mechanism:

$$
z=\operatorname{GELU}(z_f)\odot\sigma(z_g),
$$

where \(z_f\) and \(z_g\) are feature and gate channels, \(\sigma(\cdot)\) is sigmoid, and \(\odot\) is the Hadamard product. This design improves the selective representation of temporal information in the industrial robot residual.

5.3 Spatiotemporal Block Integration

Each spatiotemporal block is organized as time convolution, graph convolution, and time convolution. The first time convolution extracts local historical dynamics within each joint. The graph convolution then propagates information across joints according to the multi-relation topology. The final time convolution integrates the fused spatiotemporal features. For layer \(l\),

$$
H^{(l+1)}=
\operatorname{LN}
\left(
T_d^{(2)}
\left(
G
\left(
T_d^{(1)}(H^{(l)})
\right)
\right)
\right),
$$

where \(T_d^{(1)}\) and \(T_d^{(2)}\) are causal dilated temporal convolutions, \(G\) is the multi-relation graph convolution, and \(\operatorname{LN}\) is layer normalization. The input is first projected to 64 channels by a \(1\times1\) convolution, and then three STG blocks encode the sequence. Table 4 lists the main architectural settings.

Component Setting Value
Input window length \(W\) 20
Node feature dimension \(d_z\) 4
Hidden channel dimension \(C\) 64
Temporal kernel widths \(K\) 3 and 5
Dilation rates \(d\) 1, 2, 4
Number of STG blocks \(L\) 3
Attention scoring convolutions \(\psi\) 64 to 64 to 1
KAN grid number \(G_{\mathrm{KAN}}\) 4
KAN spline order \(k_{\mathrm{KAN}}\) 3
KAN grid range \([a,b]\) [-2, 2]

5.4 Temporal Attention Readout

After spatiotemporal encoding, I use temporal attention to focus on the most relevant parts of the window. For joint \(i\), let \(H_i\in\mathbb{R}^{C\times W}\) be its feature sequence. The attention score is

$$
s_i(t’)=\psi(H_i(:,t’)),
$$

where \(\psi\) is a scoring function composed of two convolutions and GELU nonlinearity. The normalized temporal weight is

$$
\alpha_i(t’)=
\frac{\exp(s_i(t’))}
{\sum_{\tau=t-W+1}^{t}\exp(s_i(\tau))}.
$$

The readout for joint \(i\) is then

$$
h_i^{\mathrm{att}}=
\sum_{t’=t-W+1}^{t}
\alpha_i(t’)H_i(:,t’).
$$

The joint-level readout representation is

$$
H^{\mathrm{att}}=
\begin{bmatrix}
h_1^{\mathrm{att}}, & h_2^{\mathrm{att}}, & \cdots, & h_N^{\mathrm{att}}
\end{bmatrix}.
$$

This attention readout gives the industrial robot model a dynamic summary of the recent motion history.

5.5 KAN Residual Regression Head

After readout, I concatenate the \(N\) joint features into a global vector. For \(N=6\) joints and 64 channels, the concatenated dimension is 384. The KAN regression head predicts the six-dimensional residual torque:

$$
\hat r(t)=
f_{\mathrm{reg}}
\left(
\begin{bmatrix}
h_1^{\mathrm{att}}(t), & h_2^{\mathrm{att}}(t), & \cdots, & h_N^{\mathrm{att}}(t)
\end{bmatrix}
\right).
$$

A single KAN linear unit is

$$
\operatorname{KAN}(x)=
\operatorname{SiLU}(W_b x)+W_s\Phi(x),
$$

where \(W_b\) is the base linear weight, \(W_s\) is the spline mapping weight, and \(\Phi(x)\) is the B-spline basis expansion. I use three KAN linear layers. The first two layers are followed by layer normalization. The spline grid number is 4, the B-spline order is 3, and the grid range is \([-2,2]\). The final hybrid prediction is

$$
\tau_{\mathrm{final}}(t)=
\tau_{\mathrm{pred}}(t)+\hat r(t).
$$

This hybrid form preserves the physical baseline of the industrial robot and adds a flexible nonlinear correction.

6. Training and Evaluation Protocol

I train the residual compensator with a weighted Huber loss to improve robustness against local outliers while maintaining accurate regression. Let the true residual be \(r_j(t)\) and the predicted residual be \(\hat r_j(t)\). The loss is

$$
\mathcal{L}=
\frac{1}{N}
\sum_{j=1}^{N}
w_j h_{\delta}
\left(
r_j(t)-\hat r_j(t)
\right),
$$

where \(w_j\) is the weight of joint \(j\), and \(h_{\delta}(\cdot)\) is the Huber function:

$$
h_{\delta}(e)=
\begin{cases}
\dfrac{1}{2}e^2, & |e|\le\delta,\\[6pt]
\delta\left(|e|-\dfrac{1}{2}\delta\right), & |e|>\delta.
\end{cases}
$$

I use the root-mean-square error as the main metric:

$$
\mathrm{RMSE}=
\sqrt{
\frac{1}{K}
\sum_{k=1}^{K}
\left(
\tau_{m,k}-\tau_{p,k}
\right)^2
},
$$

where \(\tau_{m,k}\) and \(\tau_{p,k}\) are measured and predicted torque at sample \(k\), and \(K\) is the number of samples. Table 5 lists the unified training settings. Each model uses the same data split, preprocessing, metric, and training budget for fair comparison.

Parameter Value Search range
Optimizer AdamW Fixed
Initial learning rate \(5\times10^{-4}\) \(2\times10^{-4}\) to \(1\times10^{-3}\)
Batch size 64 Fixed
Dropout rate 0.1 0 to 0.2
Maximum epochs 60 Fixed
Residual window 20 Fixed
Random seeds 1, 11, 21, 42 Fixed

I also compare the trainable parameter counts of several models. The proposed STG-KAN has 1.115 million parameters. The baseline BP model has 0.411 million, LSTM has 0.587 million, GRU has 0.175 million, and the ablation variants have 1.064, 0.513, and 0.711 million parameters, respectively. Table 6 summarizes these counts.

Model Trainable parameters (M)
BP 0.411
LSTM 0.587
GRU 0.175
STG-KAN 1.115
Ablation without graph module 1.064
Ablation without temporal module 0.513
Ablation with MLP head 0.711

7. Experimental Platform and Data

I evaluate the proposed hybrid model on a six-degree-of-freedom serial industrial robot. The robot is controlled and measured through a real-time automation environment at 100 Hz. Joint positions, velocities, and drive torques are recorded synchronously. Position, velocity, and torque signals are filtered by an eighth-order Butterworth low-pass filter with a 3 Hz cutoff and zero-phase filtering. Joint acceleration is estimated offline by combining Kalman filtering and RTS smoothing. The processed position, velocity, acceleration, and torque signals are used for parameter identification and residual sample construction.

The excitation trajectory covers a wide range of the workspace. I collect 60 trajectory groups. For baseline identification, one preprocessed excitation trajectory is used to construct the observation matrix and estimate the physically feasible parameters. The baseline model then predicts all trajectory data, and the residual samples are constructed. I split the data at the trajectory level: 48 trajectories for training and 12 for testing. All time windows generated from the same trajectory belong to the same subset, which prevents data leakage across adjacent windows.

The kinematic parameters of the industrial robot are listed in Table 7 in a modified DH convention. These parameters define the nominal geometry used in the baseline model.

Joint \(i\) \(\theta_i\) \(\alpha_{i-1}\) \(a_{i-1}\) (m) \(d_i\) (m)
1 \(\theta_1\) \(180^\circ\) 0 -0.500
2 \(\theta_2-90^\circ\) \(90^\circ\) 0.200 0
3 \(\theta_3\) \(0^\circ\) 0.790 0
4 \(\theta_4+180^\circ\) \(90^\circ\) 0.140 -0.780
5 \(\theta_5+180^\circ\) \(90^\circ\) 0 0
6 \(\theta_6+180^\circ\) \(90^\circ\) 0 -0.124

8. Baseline Physical Model Results

I use a fifth-order Fourier series with a 20 s duration to excite the industrial robot. The identification data are collected under this trajectory, and the physically feasible parameters are estimated by the constrained semidefinite program. The baseline model captures the overall torque trend but shows visible errors near reversals, peaks, and fast changes. Table 8 reports the baseline RMSE for each joint.

Joint \(i\) 1 2 3 4 5 6
RMSE (N·m) 21.28 38.22 26.57 8.01 3.83 1.27

For the first three joints, the baseline RMSE is within about \(15\%\) of the maximum absolute torque. For the last three joints, the baseline RMSE is within about \(20\%\) of the maximum absolute torque. These results confirm that the physical model is useful but not sufficient for high-accuracy compensation of the industrial robot.

9. Hybrid Model Comparison

I compare the proposed STG-KAN model with the baseline physical model and three deep-learning compensation models: BP, LSTM, and GRU. On a representative test trajectory, all hybrid models improve over the baseline. The proposed model achieves the lowest overall RMSE. Table 9 reports the joint-level and overall comparison.

Joint \(i\) Baseline BP LSTM GRU STG-KAN
1 31.55 21.31 22.54 24.82 13.10
2 51.14 38.04 36.74 40.59 30.08
3 23.67 18.59 18.46 19.94 14.18
4 6.52 4.90 5.76 4.52 2.46
5 5.00 3.21 3.17 3.54 2.28
6 1.45 1.03 1.05 1.13 0.84
Overall RMSE 26.59 19.50 19.33 21.20 14.66

On this representative trajectory, the proposed STG-KAN model reduces the overall RMSE from \(26.59\,\mathrm{N\cdot m}\) to \(14.66\,\mathrm{N\cdot m}\), a reduction of \(44.9\%\). Compared with BP, LSTM, and GRU, the reductions are \(24.8\%\), \(24.2\%\), and \(30.8\%\), respectively. The proposed model achieves the lowest error for all six joints, which indicates that the improvement is not limited to one joint but is distributed across the industrial robot.

To reduce the effect of random data splitting, I repeat the trajectory-level experiment four times with random seeds 1, 11, 21, and 42. Table 10 reports the mean and standard deviation of the RMSE. The proposed STG-KAN model achieves \(15.25\pm1.28\,\mathrm{N\cdot m}\). Compared with the baseline, BP, GRU, and LSTM, the reductions are \(33.75\%\), \(18.45\%\), \(23.02\%\), and \(24.47\%\), respectively.

Model RMSE mean ± standard deviation (N·m) Reduction relative to baseline
Baseline physical model \(23.02\pm0.91\) —
BP \(18.70\pm0.96\) 18.77%
GRU \(19.81\pm0.76\) 13.94%
LSTM \(20.19\pm0.83\) 12.29%
STG-KAN \(15.25\pm1.28\) 33.75%

The BP model improves the baseline to some extent, but it lacks explicit temporal and topological modeling. LSTM and GRU use historical information, but they do not explicitly represent the serial topology of the industrial robot. The proposed STG-KAN model jointly represents directional inter-joint coupling and historical hysteresis, which explains its lower prediction error and better residual compensation.

10. Ablation Study

I conduct ablation studies to evaluate the contribution of the spatial graph module, the temporal module, and the KAN regression head. Three variants are considered. The first replaces the multi-relation graph convolution with a fully connected layer. The second removes the multiscale temporal encoding. The third replaces the KAN head with an ordinary MLP head. All other training and testing settings remain identical. Table 11 reports the results over four repeated trajectory-level experiments.

Model RMSE mean ± standard deviation (N·m) Increase relative to full model
STG-KAN \(15.25\pm1.28\) —
Case 1: without graph module \(17.76\pm1.62\) 14.13%
Case 2: without temporal module \(18.94\pm1.14\) 19.48%
Case 3: with MLP head \(17.53\pm1.38\) 13.01%

The full model achieves the lowest RMSE. Removing the graph module increases the error by \(14.13\%\), which shows that explicit topology modeling is important for the industrial robot residual. Removing the temporal module increases the error by \(19.48\%\), which is the largest degradation among the three variants. This indicates that multiscale temporal encoding is critical for friction reversal and hysteresis. Replacing the KAN head with an MLP head increases the error by \(13.01\%\), which confirms that the KAN head improves nonlinear residual fitting. Table 12 summarizes the qualitative role of each module.

Module Function Observed effect
Multi-relation graph convolution Models directional joint coupling and nonlocal interaction Removing it increases RMSE by 14.13%
Multiscale causal temporal encoding Models short-term fluctuation and long-term hysteresis Removing it increases RMSE by 19.48%
Temporal attention readout Selects informative historical segments Improves temporal aggregation
KAN regression head Improves nonlinear residual approximation Replacing it increases RMSE by 13.01%

11. Discussion

The experimental results show that residual compensation is necessary for an accurate inverse dynamics model of an industrial robot. The baseline physical model provides a stable and interpretable prediction, but it cannot represent all friction, compliance, and coupling effects. The proposed hybrid model keeps the physical baseline and learns only the residual, which reduces the burden on the neural network and improves generalization. This is important for an industrial robot because pure black-box models may fail outside the training distribution or produce physically inconsistent torque predictions.

The multi-relation graph is a key component. In a serial industrial robot, the effect of joint \(i\) on joint \(j\) is not symmetric. Downward propagation and upward feedback have different dynamic meanings. By using separate relation graphs, the model can distinguish these directions. The learned correction \(\Delta A\) further allows the network to discover nonlocal coupling patterns that are not captured by a simple distance-decay prior. This is particularly useful for an industrial robot carrying a payload or experiencing configuration-dependent inertial coupling.

The temporal module is also essential. Friction reversal near zero velocity produces sharp residual changes, while actuator and transmission hysteresis produce longer memory effects. A single-scale temporal convolution cannot capture both. The multiscale causal dilated convolution with gated fusion allows the model to represent both short and long dependencies. The temporal attention readout then emphasizes the most relevant samples inside the window. For an industrial robot performing repeated or quasi-periodic motions, this mechanism improves the compensation of direction-dependent errors.

The KAN regression head provides additional flexibility. An ordinary MLP head can approximate nonlinear functions, but it may require more parameters or wider layers for highly nonlinear residual surfaces. The KAN head uses spline basis expansion, which gives a different inductive bias and improves local nonlinear fitting. In the ablation study, replacing KAN with MLP increases the RMSE by \(13.01\%\), which supports this design choice for the industrial robot residual compensator.

12. Practical Implications for Industrial Robot Control

The proposed hybrid model can be used as a feedforward torque predictor for an industrial robot. In a model-based controller, the predicted torque can reduce tracking error and improve disturbance rejection. Because the baseline model remains physically interpretable, the compensation term can be monitored separately. If the residual becomes unusually large, this may indicate payload change, friction degradation, or a fault. Therefore, the hybrid model can support both control and condition monitoring of an industrial robot.

The computational cost is also acceptable. The STG-KAN model has 1.115 million parameters, which is larger than BP, LSTM, and GRU, but still practical for offline training and real-time inference on modern industrial controllers. The causal convolution and sliding-window design allow streaming implementation. The graph structure is fixed by the serial topology of the industrial robot, so the adjacency matrices can be precomputed. The temporal convolution and attention operations are parallelizable over the window, which further supports real-time deployment.

Table 13 summarizes the practical benefits and limitations of the proposed method.

Aspect Benefit Limitation
Physical baseline Interpretable and stable prediction Cannot model all friction and compliance effects
Residual compensation Improves local and global torque accuracy Depends on training data coverage
Multi-relation graph Captures directional coupling of industrial robot joints Requires fixed joint topology
Temporal convolution Captures hysteresis and friction reversal Window length must be selected
KAN head Improves nonlinear residual fitting Adds parameters and training cost

13. Conclusion

I have presented a hybrid dynamics modeling method for an industrial robot based on a physically feasible baseline and a spatiotemporal graph neural network residual compensator with a KAN regression head. The baseline model is identified with constrained optimization and provides an interpretable physical prediction. The STG-KAN model then compensates the unmodeled residual by combining multi-relation graph convolution, multiscale causal temporal encoding, temporal attention, and KAN regression. The multi-relation graph represents downward propagation, upward feedback, and nonlocal coupling in the serial industrial robot. The temporal module represents friction reversal, actuator hysteresis, and historical dependence. The KAN head improves nonlinear residual approximation.

Across four trajectory-level repeated experiments, the proposed model achieves an overall RMSE of \(15.25\pm1.28\,\mathrm{N\cdot m}\), which is \(33.75\%\) lower than the baseline physical model. It also outperforms BP, LSTM, and GRU compensation models. Ablation studies confirm that the graph module, temporal module, and KAN head each contribute to the final accuracy. The results suggest that combining physical feasibility with spatiotemporal residual learning is an effective way to improve the inverse dynamics prediction of an industrial robot while retaining interpretability.

In future work, I plan to extend the hybrid framework to payload adaptation, friction identification under varying temperatures, and real-time feedforward control of an industrial robot. I also intend to investigate uncertainty estimation for the residual compensator so that the industrial robot controller can adjust its trust in the learned compensation according to operating conditions.

Scroll to Top