The Shape of Learning: How Robot Arms Get Smarter Depends on Their Activation Function
The Robot's New Trick: Teaching Arms to Feel Their Way Through Space
In a robotics lab in Turku, Finland, something small but significant is happening. A mechanical arm—a Quanser QArm, to be precise—is learning to track a moving target with uncanny precision. It follows a smooth sine wave, pivoting through space to keep pace with an invisible partner. The tracking error, the distance between where the arm points and where it should point, hovers around a fraction of a degree. When the target changes direction abruptly—from a flowing curve to a sharp square—the arm adjusts, corrects, and settles, sometimes with a slight overshoot that the system's neural brain quickly damps out.
What makes this noteworthy isn't the arm itself. Robotic manipulators have been performing tasks like this for decades. What's remarkable is how the arm achieves its precision, and specifically, how one particular choice—hidden in the mathematical heart of its control system—can mean the difference between good tracking and great tracking.
The choice is called the activation function, and the researchers who built this system have just published the most comprehensive study to date on how activation functions shape the behavior of intelligent robotic controllers. Their findings, published in July 2026 by a team from the University of Turku, reveal something that many engineers have suspected but never systematically demonstrated: that the kernel shape of a neural network isn't merely an implementation detail. It's a design parameter as fundamental as choosing the right gear ratio or sensor resolution.
The team's central result is striking. When tracking smooth, continuous trajectories, a particular kernel called the inverse multiquadratic function reduced tracking error by 51.9% compared to a conventional controller—one that doesn't use neural networks at all for disturbance estimation. Control effort, meanwhile, increased by less than 3%. The implications are significant: you can dramatically improve accuracy without meaningfully increasing energy consumption.
But—and this is the nuance that makes the paper valuable—the benefits aren't universal. When the arm had to track discontinuous, step-like commands, the same kernel that excelled at smooth tracking introduced overshoot and longer settling times. A more localized kernel, the Gaussian, handled those transitions more gracefully. The researchers had discovered a genuine trade-off, one that engineers can now navigate deliberately rather than stumble through blindly.
"Activation function selection acts as a structural design parameter in intelligent control," the authors write, "directly shaping adaptation dynamics and practical closed-loop performance."
This is the kind of insight that, once articulated clearly, seems obvious in retrospect. But the robotics literature has long treated activation function choice as defaulting to Gaussian without much interrogation. This paper challenges that assumption with experimental rigor.
The Problem: Making Robots Robust to Reality
To understand why activation functions matter, it helps to appreciate the challenge that robotic control engineers face every day.
A robotic manipulator—think of the arm on a factory floor or the jointed mechanism in a surgical robot—moves through space by applying torques to its joints. The physics governing this motion are well understood: the inertia of each link, the centrifugal forces as parts spin, the ever-present pull of gravity. These relationships are encoded in equations that have existed since the 1970s, and control theorists have developed sophisticated methods to exploit them.
The dominant approach is called feedback linearization. The idea is conceptually elegant: if you know the exact dynamics of your system, you can transform the nonlinear problem of robot control into a simpler linear one. You calculate what torque to apply, the dynamics cancel out the nonlinearities, and the system responds like a well-behaved linear controller. Under ideal conditions, this works beautifully.
But ideal conditions don't exist in the real world. Every robot has parameters that differ from their nominal values—slightly heavier links, slightly different joint friction, manufacturing variations that add up. There are unmodeled dynamics: flex in cables, subtle deformations, complex friction behaviors that resist clean mathematical description. There are external disturbances: a push from a human collaborator, a change in payload, thermal effects that shift parameters over time.
These uncertainties don't just degrade performance; they can destroy the theoretical guarantees that make feedback linearization appealing. A controller that assumes perfect knowledge of the system may perform well in simulation but fail catastrophically when the physical reality diverges from the model.
Engineers have developed several strategies to address this fragility. Adaptive control methods estimate unknown parameters online, adjusting the controller as the system runs. Robust control methods design controllers that work despite uncertainties, often by sacrificing some performance to gain guarantees. And increasingly, learning-based approaches use neural networks to approximate the unknown disturbances directly, learning from data what the model cannot capture.
Each approach has trade-offs. Adaptive methods work well for structured uncertainties—parameters you know are wrong but can estimate—while neural networks can handle unstructured uncertainties: the messy, complex, hard-to-characterize effects that don't fit neat parameter estimates. But neural networks bring their own challenges: they need training, they can behave unpredictably, and integrating them with model-based controllers in a way that preserves stability is far from trivial.
The paper by Paldanius, Da Silva Lima, and Bessa sits at this intersection, combining the best of both worlds. They use feedback linearization for the nominal system, exploiting what they know, while deploying a radial basis function (RBF) neural network to estimate and compensate for whatever the model gets wrong. The stability of the closed-loop system is proven formally using Lyapunov analysis—a mathematical framework that ensures all signals remain bounded and that tracking error converges to a small region around zero.
This hybrid architecture isn't new. But the systematic investigation of how the neural network's internal structure affects performance is.
The Architecture: Where Mathematics Meets Movement
The control framework developed by the researchers has three interconnected layers, each building on the last.
At the foundation is the manipulator's equations of motion, expressed in the standard form for robotic systems:
This compact notation encodes a great deal of physics. $\mathbf{M}(\mathbf{q})$ is the inertia matrix, capturing how the mass distribution of the robot affects its acceleration. $\mathbf{k}(\mathbf{q}, \dot{\mathbf{q}})$ represents the Coriolis and centrifugal forces that arise when parts of the robot move relative to each other. $\mathbf{h}(\mathbf{q}, \dot{\mathbf{q}})$ is the vector of external forces, most importantly gravity. And $\mathbf{u}$ is the control input—the torques applied to each joint.
For control design, the researchers rewrite this equation to separate the known dynamics from the unknown disturbances:
Here, $\hat{\mathbf{M}}$ and $\hat{\mathbf{f}}$ are estimates of the true inertia matrix and the combined nonlinear dynamics. The term $\mathbf{d}$ captures everything these estimates get wrong—parametric uncertainties, friction, unmodeled effects, external disturbances. This is the quantity the neural network will learn to estimate and cancel.
The control law combines feedforward prediction with feedback correction:
The notation looks dense, but the intuition is straightforward. The term $\ddot{\mathbf{q}}_d$ is the desired acceleration—the trajectory we want to follow. The term $\mathbf{s} = \dot{\tilde{\mathbf{q}}} + \mathbf{\Lambda}\tilde{\mathbf{q}}$ is a combined error signal that blends position error and velocity error, weighted by the diagonal matrix $\mathbf{\Lambda}$. By choosing the elements of $\mathbf{\Lambda}$ appropriately, the designer can tune how aggressively the controller responds to errors.
The term $\hat{\mathbf{d}}$ is the neural network's estimate of the disturbance. If this estimate were perfect, $\mathbf{d} - \hat{\mathbf{d}} = 0$, and the closed-loop dynamics would reduce to $\dot{\mathbf{s}} + \mathbf{\Lambda}\mathbf{s} = 0$, which means the error would decay exponentially to zero.
In reality, the estimate is never perfect, but as long as it's good enough, the tracking error will converge to a small neighborhood of zero. How small depends on the approximation quality of the neural network and the magnitude of any residual errors.
The neural network itself is a radial basis function network—a particular architecture well-suited to function approximation. Each element of the disturbance estimate is computed as a weighted sum of basis functions:
The vector $\mathbf{w}_i$ contains the adjustable weights, and $\boldsymbol{\varphi}_i$ contains the activation functions. The input to these functions is $s_i$, the combined error signal for the $i$-th joint. The network has a single hidden layer, and each neuron computes its activation based on the distance between its center and the current input, normalized by a width parameter.
The key design choice the paper investigates is the form of the activation function $\kappa(\rho)$, which takes the normalized distance $\rho$ as input. Three kernels are compared:
The Gaussian kernel: $\kappa_G(\rho) = \exp(-\rho^2)$, which decays rapidly as distance from the center increases, giving it strong locality.
The Laplacian kernel: $\kappa_L(\rho) = \exp(-|\rho|)$, which also decays exponentially but more slowly than the Gaussian, producing a different shape of locality.
The inverse multiquadratic kernel: $\kappa_{IMQ}(\rho) = \frac{1}{\sqrt{1 + \rho^2}}$, which has global support—meaning it never truly decays to zero, but instead contributes to the output from every part of the input space.
These three kernels represent meaningfully different mathematical shapes. The Gaussian and Laplacian are localized: far from a neuron's center, its activation approaches zero, and that neuron has little influence on the output. The inverse multiquadratic is global: every neuron contributes something, no matter where the input falls.
This distinction turns out to matter enormously for how the network learns and adapts.
The Learning Rule: Bounded Adaptation
One of the paper's technical contributions is the learning law for updating the neural network weights. Standard gradient descent could drive weights to arbitrarily large values, potentially destabilizing the system or causing the network to behave erratically. The researchers use a projection algorithm that constrains the weights to stay within a convex region.
The learning law is:
This is a Hebbian-like rule: when the error signal $s_i$ is large, the weights adjust quickly; when the error is small, learning slows. The projection operator kicks in only when the weights approach their boundary, gently projecting them back into the admissible region rather than letting them escape.
The stability proof, developed in Section 2 of the paper, shows that under this learning law, all closed-loop signals remain bounded, and the tracking error converges to a compact region around zero. Crucially, this proof holds for any bounded activation function. The theoretical guarantee is not affected by whether you use Gaussian, Laplacian, or inverse multiquadratic kernels.
This is both good news and a provocation. Good news because it means engineers have freedom to choose the kernel that works best for their application without worrying about stability. A provocation because it raises the question: if stability doesn't depend on the kernel, what does?
The answer, which the experimental section confirms, is performance—specifically, transient behavior, steady-state accuracy, and control smoothness.
The Experiment: Three Trajectories, Three Kernels
To isolate the effect of activation function choice, the researchers kept every other parameter identical across all tests. The controller gains $\lambda_i = 30$ and $\eta_i = 10$, the projection bounds $\mu_i = 100$, the number of neurons $n_i = 6$, the desired trajectories, and the simulation environment were all held constant. Only the kernel was varied.
The testbed was a digital twin of the Quanser QArm, a four-degree-of-freedom robotic manipulator, running in the Quanser Interactive Labs environment. This software platform accurately emulates the physical arm's dynamics, including joint friction and actuator constraints, at a hardware-timed sampling frequency of 500 Hz. Each trial ran for 62 seconds, and the manipulator was returned to a home position between tests to ensure identical starting conditions.
The choice of a digital twin rather than physical hardware is pragmatic—precise experimental conditions are easier to guarantee in simulation—but the dynamics model was calibrated against the real system, making the results physically relevant.
Three reference trajectories probed different aspects of performance:
Sinusoidal reference: A smooth, continuous trajectory that tests steady-state tracking under regular, predictable excitation.
Square reference: A step-like trajectory with discontinuous transitions that tests transient response and the controller's ability to track sudden changes.
Triangular reference: A ramp-like trajectory with constant-velocity segments that tests tracking under piecewise-linear motion.
Performance was quantified using three metrics:
- RMS error: The root-mean-square of the tracking error over the trial duration. Lower is better.
- IAE (Integral of Absolute Error): The total accumulated error. This penalizes persistent small errors as much as it penalizes large errors, making it sensitive to steady-state behavior.
- RMS control effort: The root-mean-square of the PWM signal sent to the actuator. This quantifies energy consumption and control activity.
For the square reference, overshoot and settling time were also reported, capturing how the controller handles discontinuous commands.
What They Found: The Trade-offs Emerge
The results, summarized across Tables 1-3 in the paper, reveal a clear pattern: no single kernel dominates across all scenarios. Each has a domain where it excels.
For sinusoidal tracking, the inverse multiquadratic kernel delivered the best performance, achieving a 51.9% reduction in RMS error compared to the baseline (the conventional feedback linearization controller without neural disturbance estimation). The Gaussian and Laplacian kernels also improved on the baseline, but by smaller margins. Critically, the improvement came with minimal additional control effort: the IMQ kernel's RMS control effort increased by less than 3% relative to the baseline.
The interpretation is intuitive once understood. A smooth sinusoidal trajectory generates a disturbance estimate that varies continuously in time. The IMQ kernel, with its global support, averages information across the entire input space, producing a more holistic estimate of the disturbance manifold. The localized kernels, by contrast, base their estimates on the nearest neurons, potentially missing broader patterns in the data.
For triangular tracking, the results were similar. The IMQ kernel reduced RMS error by 49.0% compared to the baseline, again with modest increases in control effort. The ramp-like segments of the triangular trajectory, like the sinusoidal trajectory, have a continuous character that favors global kernels.
For square tracking, the story changed. Here, the localized kernels showed advantages in transient performance. The square reference requires the arm to track discontinuous changes in position—the arm must accelerate quickly, overshoot slightly due to inertia, then correct back. This is exactly the scenario where aggressive adaptation can help or hurt.
The neural controllers improved IAE compared to the baseline (by up to 15.0% for the Gaussian case), but they introduced overshoot and longer settling times. The IMQ kernel, which had excelled at smooth tracking, showed overshoot of 6.86% in the square tests—meaning the arm initially moved past its target before correcting. The Gaussian kernel, with its more localized response, handled the transitions more gracefully.
This reveals a fundamental trade-off: the same property that makes IMQ good at smooth tracking—its global support, which enables broader generalization—makes it less nimble at responding to sudden changes. The localized kernels sacrifice some steady-state accuracy for better transient behavior.
The researchers summarize the implication directly: "kernel support and smoothness characteristics directly affect the disturbance approximation quality and the evolution of the adaptive weights." The activation function isn't just a box to check; it's a dial that can be tuned for the application's needs.
Relative Tracking Performance on Sinusoidal Reference
| Label | Value |
|---|---|
| Baseline | 100 |
| Gaussian | 65 |
| Laplacian | 55 |
| Inverse Multiquadratic | 48.1 |
Why This Changes Things: Beyond Default Assumptions
The robotics and control communities have long defaulted to Gaussian kernels in RBF networks, often without systematic justification. This paper challenges that default, demonstrating both why the default exists (Gaussian kernels are mathematically convenient, well-understood, and stable) and why it shouldn't be universal.
The practical implications are significant. Consider a few scenarios:
Precision manufacturing: A robotic arm performing polishing or painting tasks follows smooth, continuous paths. The IMQ kernel's superior steady-state accuracy would directly translate to better product quality with minimal energy cost. A 52% reduction in tracking error at the cost of 3% more control effort is an excellent trade.
Collaborative robotics: A robot working alongside humans needs to respond quickly to sudden commands—grabbing an object, avoiding a collision. The localized kernels' better transient response would serve this application better, even if steady-state accuracy is slightly lower.
Surgical robotics: Precision is paramount, and trajectories are typically smooth. IMQ's advantages compound: not only is tracking better, but the minimal increase in control effort means less mechanical wear and tear on sensitive equipment.
Rehabilitation devices: Exoskeletons helping patients regain movement need to track human-generated commands that may be irregular. The Gaussian kernel's balanced behavior might be most appropriate, providing decent steady-state performance without aggressive transients that could startle or injure patients.
These scenarios illustrate that the "best" kernel depends entirely on context. The contribution of this paper is not to declare a winner but to provide a framework for making the choice deliberately.
There are also implications for how we design intelligent controllers more broadly. The paper's framing of activation function choice as a "structural design parameter" invites broader thinking about what else in a neural network might be similarly under-specified by theory but consequential in practice. Center placement, width selection, number of neurons—these are all choices that stability proofs leave open and that engineers typically set by heuristics or brute-force search.
This paper demonstrates that systematic investigation of such choices can yield meaningful performance improvements. It sets a methodological template: keep everything else constant, vary one parameter, and measure rigorously. The experimental protocol—identical conditions, multiple trajectory types, multiple metrics—provides a model for how to conduct such investigations credibly.
What's Next: From Single Arm to Many, From Heuristics to Design Rules
The paper is explicit about its limitations and about the directions for future work.
First, the experiments were conducted on a single degree-of-freedom of the Quanser QArm, not the full four-DOF system. Real-world robotic manipulators have coupled dynamics—moving one joint affects the forces on others—and how activation function choice interacts with coupling effects remains an open question. The authors plan to extend the analysis to multi-DOF configurations, which will introduce additional complexity: each joint could in principle use a different kernel, and the interaction between kernels across coupled dynamics could produce emergent behaviors.
Second, the kernel parameters—centers and widths—were kept identical across all tests to enable a fair comparison of kernel types. In practice, these parameters also affect performance, and the interaction between kernel type, center placement, and width selection is not well understood. Systematic tuning strategies that balance approximation capability against transient robustness could unlock further improvements.
Third, the study focused on a specific class of RBF networks with a single hidden layer. More complex architectures—deep networks, recurrent networks, networks with more layers—might exhibit different sensitivities to activation function choice. The stability proofs in the paper rely on properties specific to single-layer RBF networks, and extending them to deeper architectures is a non-trivial theoretical challenge.
There are also practical questions about implementation. The paper uses digital twin experiments, which provide precise control over conditions but may not capture all the phenomena present in physical hardware. Thermal effects, mechanical wear, sensor noise, and electromagnetic interference could affect which kernel performs best in real deployments. Experimental validation on physical hardware would strengthen the conclusions.
More broadly, the paper opens a window onto a research agenda: using systematic experimental investigation to guide neural network design in control applications. The control theory literature has traditionally focused on stability and convergence, treating performance as a secondary concern to be addressed later or via heuristics. This paper demonstrates the value of inverting that priority—starting with performance as the primary concern and recognizing that stability proofs are necessary but not sufficient for practical success.
The findings also invite broader reflection on the nature of intelligent control. The word "intelligent" appears in the paper's title, and the neural network's ability to learn the disturbance dynamics is central to the approach. But the specific results show that intelligence here is not a property of the network alone; it's a property of the network in context, shaped by the geometry of its activation functions. This is a reminder that what we call intelligence in machines is always a product of architecture, learning rules, and domain characteristics working together.
Relative Tracking Performance on Triangular Reference
| Label | Value |
|---|---|
| Baseline | 100 |
| Gaussian | 85 |
| Laplacian | 78 |
| Inverse Multiquadratic | 51 |
The Broader Significance: Precision, Robustness, and the Future of Robotics
Robotic manipulators are increasingly ubiquitous: in factories, hospitals, warehouses, and homes. Improving their precision, robustness, and efficiency has tangible consequences for productivity, safety, and quality of life. The control algorithms that govern these machines are often invisible to end users, but they determine everything from how smoothly a robot moves to how much energy it consumes.
The traditional view of robotic control has emphasized model accuracy: if you know the dynamics well enough, you can compensate for everything. This view has produced elegant theories and effective controllers, but it has also led to brittle systems that fail when models are imperfect—which is to say, always.
The intelligent control paradigm takes a different approach: acknowledge that models are imperfect, build systems that learn from experience, and ensure that learning doesn't destabilize the closed loop. This paper advances that paradigm by showing that the learning machinery itself—specifically, the shape of the neural network's activation functions—has a systematic effect on closed-loop performance that can be exploited by designers.
The result is a more nuanced design space. Engineers no longer need to choose between model-based precision and learning-based adaptability. They can have both, but they must make thoughtful choices about how to combine them. This paper provides one of the most thorough explorations yet of how to make those choices well.
The practical message is clear: activation functions are not implementation details but design parameters, and different applications call for different choices. The inverse multiquadratic kernel excels at smooth, continuous trajectories where steady-state accuracy matters most. The Gaussian kernel provides balanced performance suitable for general-purpose applications. The Laplacian kernel occupies intermediate territory.
These are not just abstract recommendations; they are grounded in experimental data from a physically relevant testbed. The 52% error reduction achievable with IMQ on sinusoidal trajectories, at a cost of less than 3% additional control effort, represents a meaningful improvement that could be realized in practice.
As robotic systems become more sophisticated and deployed in more demanding environments, the insights in this paper become increasingly valuable. The next time an engineer designs an intelligent controller for a robotic manipulator, they will need to make choices about activation functions. This paper ensures those choices can be made deliberately, with clear understanding of the trade-offs involved.
The arm in Turku may be small, but the questions it raises are large. How do we build machines that learn and adapt without losing stability? How do we choose the right learning architecture for the task at hand? How do we balance the demands of transient responsiveness against steady-state precision?
This paper doesn't answer all these questions, but it takes them seriously and provides rigorous experimental evidence to guide future work. In the end, that is what good science does: it clarifies the design space, identifies the trade-offs, and leaves the community better equipped to make good choices.
The activation function is no longer just a kernel shape in a neural network. It's a dial that engineers can tune—and now, thanks to this research, they know what each setting does.
Summary
This paper investigates how the choice of activation function in a radial basis function neural network affects the performance of an intelligent controller for robotic manipulators. Using a hybrid architecture that combines feedback linearization with online neural disturbance estimation, the researchers systematically compared Gaussian, Laplacian, and inverse multiquadratic kernels across sinusoidal, square, and triangular reference trajectories.
The key findings reveal a meaningful trade-off: the inverse multiquadratic kernel, with its global support, achieved 51.9% reduction in tracking error for smooth sinusoidal trajectories with less than 3% increase in control effort. However, the same kernel introduced up to 6.86% overshoot for discontinuous square trajectories, while the more localized Gaussian kernel handled transients more gracefully.
The results demonstrate that activation function selection is a structural design parameter in intelligent control, not merely an implementation detail. Different kernels suit different applications: global kernels excel for smooth precision tasks, while localized kernels provide better transient response for applications requiring rapid direction changes. The paper provides experimental evidence for these trade-offs and establishes a methodological template for systematically investigating neural network design choices in control applications.
The implications extend to any application where robotic manipulators must track trajectories with accuracy and robustness: precision manufacturing, surgical robotics, collaborative automation, and rehabilitation devices all stand to benefit from deliberate kernel selection informed by this research.