Many robot controllers use a neural network as a learned map from joint angles to positions, and then read the controller's command off the network's derivative, the Jacobian. A Jacobian is the matrix that says how a small joint motion moves the output. In that setting the network's derivatives matter more than its predicted values. Smooth activation functions are known to help here. Analyzing Neural Jacobian Methods in Applications of Visual Servoing and Kinematic Control showed that a Tanh network, differentiated automatically, gives smooth commands on a 7-joint arm, while a ReLU network gives jerky ones. That result covers one open-chain arm. Two questions remain open. Does the advantage hold for mechanisms whose links form closed loops, where derivatives misbehave near singular "toggle" poses? And can we choose a good network for such a mechanism without collecting much data on it?
We study both questions on a concrete mechanism from fluid machinery: a cage of guide vanes around a cross-flow wind turbine, with all vanes pitched together by a single closed-loop ring linkage. The gap between adjacent vanes sets the throat area, and mass conservation turns that area into flow acceleration. Controlling the throat area therefore means controlling a closed-chain mechanism through its Jacobian. Our claim has two parts. First, rankings of candidate networks transfer from cheap open-chain proxy mechanisms to closed-loop targets only when the networks are scored on derivative error. Second, activation smoothness is the architectural choice that drives that transfer.
Core Claim
Work on architecture transfer finds that rankings of candidate networks stay stable across datasets. NAS-Bench-201 shows this across CIFAR and ImageNet-16-120, On Network Design Spaces for Visual Recognition shows it from small proxies to ImageNet, and SMASH moves connectivity patterns from CIFAR-100 to ModelNet10. All of these score networks by classification error. The open problems in this line of work admit that rankings may collapse when the loss is physical or stiff. We predict a specific version of that collapse. If candidates for a kinematic surrogate are ranked by how well they predict output values, the ranking will not carry over to closed-loop mechanisms. If they are ranked by how well their autodiff Jacobians match the true ones, the ranking will carry over.
The mechanism behind this prediction is simple. A ReLU network can fit values well with enough width. Its Jacobian, however, is piecewise constant, and its second derivative is zero almost everywhere. Closed-loop linkages near toggle poses need accurate first and second derivatives, so smoothness of the activation decides how well a network performs there. Smoothness is also a property that does not depend on which mechanism the network is fit to. Two theory papers make this plausible: Global Convergence in Neural ODEs needs Lipschitz-smooth, non-polynomial activations for well-posed derivative dynamics, and The Loss Surface of Deep and Wide Neural Networks ties analytic activations to non-degenerate critical points. Both motivate the claim, but neither tests it on mechanisms.
The Application: Throat Area as a Kinematic Output
Guide vanes can more than double a cross-flow turbine's power coefficient once their pitch is tuned (411 Experimental Study on Optimized Guide Vanes). Fixed vanes and diffusers, though, are tuned for a narrow band of inflow conditions (OS2-11 nosecone and diffuser casing). Varying the pitch with a linkage is the natural fix, and it turns a flow-redirection problem into a kinematics problem. Let $\theta$ be the actuator angle that drives the ring and $A_t(\theta)$ the throat area between adjacent vanes. For incompressible flow through a passage of inlet area $A_{in}$, continuity gives a throat-to-inlet velocity ratio of
$$\frac{v_t}{v_{in}} = \frac{A_{in}}{A_t(\theta)},$$
and the controller tracks a target area trajectory $A^*(t)$ with $\dot\theta = \hat J(\theta)^{+}\dot A^*$. Here $\hat J = \partial \hat A_t/\partial\theta$ is the derivative of the learned surrogate $\hat A_t$, and $^{+}$ denotes the pseudo-inverse.
We use flow only as the application setting and through this geometric relation. We do not model aerodynamics with neural networks and make no claims about turbine power. In the real device, the learned surrogate is needed because the as-built linkage differs from its nominal geometry through tolerances, backlash, and deflection under aerodynamic load.
Research Program
We build one population of about 200 multilayer perceptrons. They vary in depth, width, plain versus residual connections, and activation: ReLU, Softplus with sharpness $\beta$ swept from 1 to 100 (as $\beta$ grows it approaches ReLU), Tanh, SiLU, GELU, and sine. Each network learns a forward-kinematics map from inputs to outputs, and its Jacobian is obtained by automatic differentiation.
Each network is trained on two families of mechanisms. The proxy family consists of open chains with exact simulated ground truth: planar multi-link chains and the Digit, UR5 and Panda arms used by Neural Inverse Kinematic, plus a simulated Kinova Gen-3 matching the Neural Jacobian study. The target family consists of mechanisms that break the single-joint, open-chain assumption. It includes multi-link kinematic trees taken from PartNet-Mobility, the dataset used by FlowBot++. It also includes closed-loop linkages that we solve numerically by loop closure: a four-bar, a Stephenson six-bar, and the guide-vane ring linkage. For the ring linkage the inputs are the actuator angle plus up to three perturbed geometry parameters, and the output is $A_t$.
For each network we compute two proxy scores, value error and Jacobian error, each averaged over the proxy mechanisms. We then measure four quantities on each target. The first is Jacobian error, with configurations stratified by their distance to the nearest singular pose of the linkage. The second is error in the Jacobian's derivative. The third is closed-loop area-tracking error under the control law above. The fourth is command roughness, measured as the total variation of $\dot\theta$.
The core analysis is the Kendall rank correlation between each proxy score and each target metric. We run it in two ways: across the whole population, and within groups that share an activation. The within-group version separates two effects: rankings that transfer because of smoothness, and rankings that transfer because of capacity.
Baselines
The naive reference is a ReLU network of the same size, trained on the same data. We compute its Jacobian two ways, by autodiff and by central finite differences. This is the default practice. It also tells us whether any smooth activation is doing more than finite differences already do for free.
The strongest learned competitor is Bi-NJ from Analyzing Neural Jacobian Methods. Bi-NJ learns the Jacobian directly with a secant loss and a penalty on the condition number, so the activation never has to be differentiated. That makes it the method most likely to make our smoothness story irrelevant. Its reported result is a 15.6% gain in multi-point tracking success rate over the best non-neural baseline on the Kinova Gen-3 setup. On closed-loop linkages, no reported result was found on Lacuna for Bi-NJ or for any learned surrogate. On the guide-vane linkage we will run Bi-NJ with the same training points, the same parameter budget, and the same hyperparameter search budget of 50 trials per method. Our method beats it if, on the closed-loop targets near singular poses, the best proxy-selected smooth autodiff network matches or exceeds Bi-NJ's area-tracking error and command roughness without the extra condition-number objective.
The third baseline is the analytic loop-closure model with its geometry calibrated by least squares on the same target samples. When the mechanism is fully known, this is the physically correct answer. A learned surrogate must beat it once we inject unmodelled deviations, such as backlash and load-dependent deflection, or it has no reason to exist.
The ablation that isolates the key ingredient is the Softplus sharpness sweep, with architecture held fixed. Raising $\beta$ toward ReLU-like behavior should make Jacobian transfer degrade monotonically if smoothness is the cause.
Why not just…?
Why not just use the analytic linkage model? For an ideal linkage, the analytic model is exact and nothing learned will beat it. That is why the calibrated analytic model is a baseline rather than a straw man. Our argument holds only in the regime where the as-built mechanism departs from any model with a few parameters. We will measure how large that departure must be before learned surrogates win. If the calibrated model wins at every realistic level of unmodelled error, we will report that the guide-vane case does not need learning, and the transfer findings will stand on the other target mechanisms alone.
Why not just learn the Jacobian directly, as Bi-NJ does? This is the most serious threat. Learning $J$ as an output sidesteps differentiation of the activation entirely. It has two weaknesses. It does not provide a consistent value function $A_t$, and a separately learned $J$ need not integrate to any area function. It also gives no second derivatives without a further network. Our head-to-head comparison near toggle poses settles which approach is better. If Bi-NJ wins there, the right conclusion is that derivatives should be supervised directly, and we will say so.
Why not just tune the architecture on the target mechanism? Training small networks is cheap, so compute is not the argument; the cost is target data. On a physical vane rig, every labelled point needs a measurement. We will compare proxy-selected networks against networks tuned on target validation sets of 100, 1,000 and 10,000 points. Proxy selection is useful only if it beats target tuning at the smaller budgets. IKNet reports that fine-tuning on 1,000 samples matches training from scratch on 100,000, which supports the premise that target data is the scarce resource.
Decisive Checks
The proposal succeeds if three things hold on the closed-loop targets. Rank correlation using the Jacobian-error proxy should clearly exceed correlation using the value-error proxy. That gap should largely vanish when the comparison is made within an activation group. Degradation along the Softplus sharpness sweep should be concentrated near singular poses.
The proposal fails if any of three things happens. Value-error and Jacobian-error proxies may rank networks equally well. ReLU with finite differences may match the smooth networks. Bi-NJ may dominate near singular poses. We will report whichever outcome occurs.
As a stretch goal, if a sub-scale vane rig is available, we will check that tracking the commanded throat area reproduces the velocity ratio predicted by continuity, measured with a pitot probe. This check is a sanity test of the application setting, not a test of the central claim.
References
- Analyzing Neural Jacobian Methods in Applications of Visual Servoing and Kinematic Control
- NAS-Bench-201: Extending the Scope of Reproducible Neural Architecture Search
- On Network Design Spaces for Visual Recognition
- SMASH: One-Shot Model Architecture Search through HyperNetworks
- Global Convergence in Neural ODEs: Impact of Activation Functions
- The Loss Surface of Deep and Wide Neural Networks
- 411 Experimental Study on Optimized Guide Vanes of Vertical Axis Type of Cross Flow Wind Turbines
- OS2-11 Performance improvement of micro wind turbine with a combination of nosecone and diffuser casing
- Neural Inverse Kinematic
- FlowBot++: Learning Generalized Articulated Objects Manipulation via Articulation Projection