A robot hand holds an object most securely when its fingers wrap around it. Grasping researchers measure this with the wrap-up rate, the angular fraction of the object that a multi-jointed finger encloses between its first and last contact. A classical rule of thumb treats a wrap-up rate above 0.8 as stable. The argument here is that wrapping has a hidden cost once the fingers react to touch through a feedback loop with any delay. Each new finger segment that touches the object adds another corrective force acting on the same body, so the total feedback strength of the loop grows as the hand closes. A touch-reactive controller tuned at the first fingertip contact can therefore become oscillatory by the time the hand is fully wrapped. The fingers chatter: they bounce off and re-engage the surface, and they can eject light objects or crush soft ones.
Computational neuroscience has studied the same failure as the overregulation catastrophe. In networks that cancel errors through delayed negative feedback, raising the feedback strength improves precision only up to a critical value, and beyond it the network falls into delay-driven oscillation. We propose to carry that theory into multi-link grasping. We will measure how the critical feedback gain shrinks as envelopment grows, and we will test whether scheduling each link's gain from a cheap running estimate of envelopment, together with breaking the lockstep timing of the link controllers, gives fast, chatter-free power grasps under the 5–40 ms latencies of real tactile pipelines.
Core Claim
The claim has three parts, ordered from most to least certain. First, in a reactive power grasp the stability boundary is not a fixed property of the controller. It moves with the grasp configuration, and the critical per-link gain falls roughly in inverse proportion to the strongest coupling among the engaged contacts, which tracks the wrap-up rate. Second, the precision-versus-stability trade-off follows the shape that balanced-network theory predicts. Force-holding error first falls and then rises with gain, and the best gain sits at a roughly fixed fraction of the moving critical gain. In the rate networks of Predictive coding in balanced neural networks with noise, chaos and delays, that fraction is one half, and the minimum error grows with delay. Third, and most speculative: making or breaking contact is a threshold event, much like a neuron firing. Identical delays across links should therefore produce synchronized overcorrection, the analogue of the "blind window" synchrony described in Optimal noise level for coding with tightly balanced networks of spiking neurons in the presence of transmission delays. Desynchronizing the links, either by staggering their update phases or by adding small calibrated dither, should then raise the critical gain.
The mechanism can be written with one standard model. Let $e_i$ be the force error at contact $i$. Let each link correct its own error from a measurement that is $d$ seconds old, with per-link gain $g$. Let the object couple the contacts through a matrix $A$ that depends on contact normals and contact stiffness $\kappa$. Linearized, the dynamics are
$$\dot{e}(t) = -g\,\kappa\,A\,e(t-d).$$
Every mode of a first-order delayed system $\dot{y} = -\lambda y(t-d)$ is stable only when $\lambda d < \pi/2$. This is a textbook result for scalar delay equations. It gives a critical gain
$$g_c = \frac{\pi}{2\,d\,\kappa\,\lambda_{\max}(A)},$$
where $\lambda_{\max}(A)$ is the largest eigenvalue of the coupling matrix. Adding contacts that push on the object along overlapping directions raises $\lambda_{\max}$. This is the formal sense in which envelopment shrinks the stability margin. The linear model is the hypothesis, not the result: real contacts are unilateral and frictional, and the open question is whether this simple boundary survives them.
Research Program
The work builds on four strands, each playing a different role. Envelopment geometry comes from Appropriate Lengths Between Phalanges of Multijointed Fingers for Stable Grasping, which defined the wrap-up rate in planar simulation of 28 finger designs. It found that designs with a long middle segment ("M-type") wrap better than human-like tapering ("D-type") designs. That analysis is purely kinematic and has no feedback. We add the dynamics and ask whether the best-wrapping designs pay for it with the narrowest stability margin.
The motivating gap comes from High-Bandwidth Tactile-Reactive Control for Grasp Adjustment, which runs fingertip-only tactile corrections at 200 Hz. It has no model of latency, and its authors note that aggressive corrections displace light objects and that extending the method to multi-segment fingers is beyond its scope. Two dexterous-hand systems motivate the problem from the opposite side. RESPRECT closes the hand essentially open-loop at contact. DECO executes multi-step action chunks from a slow diffusion policy. Both avoid delayed-feedback instability by not reacting during contact, and they pay for it in disturbance rejection.
The theory comes from the balanced-network papers already cited. Bifurcation work on coupled oscillators, Precise Self-tuning of Spiking Patterns in Coupled Neuronal Oscillators, adds a warning: pushing adaptation gain too high can also cause quasi-periodic drift, chaos, or complete silencing. In a hand, silencing corresponds to fingers backing off the object entirely. We will look for all three failure modes, not only oscillation.
Two papers constrain the design. Toward Practical Equilibrium Propagation showed that deliberately weak feedback speeds settling in recurrent networks. This suggests the best gain may sit well below the margin, and our gain sweeps test that directly. The neuromorphic swimming circuit of Patel et al. showed that over-strong coupling collapses coordinated segments into lockstep synchrony. That is the hardware precedent for our desynchronization hypothesis.
Concretely, we will build four things.
The first is a planar simulator of three-link fingers closing on Okada's circles and rectangles, with configurable sensing delay and compliant unilateral contacts.
The second is a 3D MuJoCo setup with a multi-fingered hand and YCB objects. These are the same 47 objects used by PointNetGPD. The setup will include transport-acceleration disturbances.
The third is an online estimator of envelopment. It computes the wrap-up rate from joint angles and tactile contact locations. It also computes a geometric proxy for $\lambda_{\max}(A)$ from contact normals alone, with $\kappa$ replaced by a crude constant, so the scheduler needs no object model.
The fourth is the scheduled controller. It sets each link's gain to a fraction $\rho$ of the estimated critical gain, and it optionally staggers link update phases or injects dither.
The fraction $\rho$ is tuned once on training objects. We then test whether one value transfers across objects, delays, and finger designs. A transferable $\rho$ is what would make this a design rule rather than a per-task tuning trick.
Baselines
The naive reference point is open-loop envelopment. The hand closes to a target posture under position control with compliant pads and no tactile reaction, as in the terminal phase of RESPRECT. It bounds what feedback buys. If it matches the reactive controllers on success and slip under transport disturbance, the problem we study does not matter in practice. Its grasp trajectories and object set are identical to everything else.
The strongest existing reactive method is the cross-finger gradient-descent controller from High-Bandwidth Tactile-Reactive Control for Grasp Adjustment. Its reported numbers are 99% simulated success on torus shapes in MuJoCo (against 2% for uncoordinated projected gradient descent) and convergence below 10° contact-normal error within 1–2 s on hardware. It is the strongest competitor because it is the only cited system that reacts to touch at high rate during contact. It is fingertip-only, so we run it in two forms. One applies it to the distal links of our hand. The other is a naive multi-link extension that applies its per-contact correction at every touching segment with one fixed gain. Its gain gets the same tuning budget as our method, a grid search on the training objects at each delay. Reported results exist only for success and alignment. For our chatter and force-oscillation metrics we found no reported result on Lacuna, so all chatter comparisons are run by us.
The third baseline is the best fixed gain. A single gain is tuned by the same grid search to maximize success over the whole closing trajectory at each delay. This is the standard engineering answer and the real competitor to scheduling.
The fourth is an oracle delay-margin scheduler. It computes $g_c$ from the true contact stiffness and true coupling matrix, information a real robot would not have. It tells us how much is lost by using the cheap geometric proxy.
The ablation that isolates the key ingredient replaces the envelopment proxy with a raw count of touching links. If scheduling by count does as well as scheduling by the normal-aware proxy or by the wrap-up rate, then contact geometry adds nothing and the honest conclusion is "scale gain by contact count." A second ablation removes desynchronization.
Beating the baselines means that, at a delay of 20 ms or more, the scheduled controller reaches a lower chatter rate and lower slip than the best fixed gain, with convergence time no worse than the fixed gain tuned for the fully wrapped configuration. It must also come within a small gap of the oracle. All comparisons use the same objects, disturbance seeds, and tuning budget.
Why not just…?
Why not just turn the gain down? A gain tuned to be safe at full wrap is overly cautious at first contact. In the balanced-network results, error falls steeply with feedback strength, so low gain should cost the most in exactly the phase where fingers still need to find the surface. This is a real threat, because the cost may be only a few hundred milliseconds of convergence. Measuring that cost is the purpose of the best-fixed-gain baseline. If the convergence penalty is small and slip is equal, scheduling is not worth its complexity, and we will report that.
Why not just run the loop faster? Latency has floors that faster buses do not remove. Optimization in the loop, camera-based tactile sensors, and diffusion-policy inference like DECO's each impose one. Our model also says the critical gain scales as $1/d$, so faster loops shift the boundary but do not remove its dependence on envelopment. The delay sweep from 5 ms (roughly 200 Hz) to 40 ms shows directly whether the effect disappears at realistic rates. If it vanishes below 10 ms, the work matters only for slow, perception-heavy pipelines, and we will say so.
Why not just use standard delay compensation, such as a predictor or classical margin analysis? Classical margin analysis is precisely what the oracle baseline does, and it needs object stiffness and contact geometry that a robot handling novel objects lacks. A Novel Simulation-Based Quality Metric for Evaluating Grasps on 3D Deformable Objects shows how expensive model-based evaluation of soft contacts is: it requires material parameters and offline FEM simulation. A predictor-based compensator has the same dependence on a plant model. Our bet is that a model-free geometric proxy recovers most of the oracle's benefit. The proxy-versus-oracle gap measures that directly.
Why not just add compliance? Soft pads lower the effective contact stiffness $\kappa$, which raises $g_c$. This may be the cheapest fix. On soft objects, though, Minimal Work and SoGraB show that the right compliance depends on the object. We include pad stiffness as a sweep axis. If one pad stiffness removes chatter across our object set, it beats our method.
Proposed Experiments and Decisive Checks
The first experiment maps the stability boundary in the planar suite. For each of Okada's five finger types and each object size, we sweep gain and delay along closing trajectories. We record the gain at which chatter begins (contact break events per second and oscillatory force power above a threshold) against the current wrap-up rate and the coupling proxy. The decisive check is whether the critical gain falls with envelopment roughly as the $1/\lambda_{\max}$ prediction says. If it does not, the core claim fails. That could happen if opposing contacts cancel instead of adding, or if friction damps the oscillation. A secondary question is whether M-type fingers, which wrap best, have the narrowest margin. That would be a concrete design trade-off absent from the kinematic analysis.
The second experiment tests the shape of the trade-off. At each envelopment level we plot holding error against gain as a fraction of $g_c$. We ask whether the optimum sits near one half, as in the balanced-network theory, or much lower, as the weak-feedback result would suggest, and how minimum error scales with delay.
The third experiment runs the full 3D comparison on YCB objects in MuJoCo. Metrics are grasp success, slip and drop under scripted transport accelerations, chatter rate, peak contact force, final wrap-up rate, and convergence time, all against the baselines above.
The fourth experiment tests desynchronization. It compares lockstep link updates, staggered phases, and dither at several amplitudes. We check whether the critical gain rises, and whether the best dither amplitude grows with delay in a way consistent with the $\delta^{1/3}$ law. We treat that law as a hypothesis, since it was derived for spiking units tracking a one-dimensional signal. A null result here leaves the first three parts of the proposal intact.
A deformable extension reuses the Isaac Gym FEM objects of Le et al. to ask whether object softness itself widens the margin.
The project succeeds if the envelopment-dependent boundary appears, and if the scheduled controller beats the best fixed gain on slip and chatter at delays of 20 ms or more without slower convergence. Its value then depends on the ablation: whether geometry-aware scheduling beats simply counting contacts.
References
- Predictive coding in balanced neural networks with noise, chaos and delays
- Optimal noise level for coding with tightly balanced networks of spiking neurons in the presence of transmission delays
- Appropriate Lengths Between Phalanges of Multijointed Fingers for Stable Grasping
- High-Bandwidth Tactile-Reactive Control for Grasp Adjustment
- RESPRECT: Speeding-up Multi-Fingered Grasping With Residual Reinforcement Learning
- DECO: Decoupled Multimodal Diffusion Transformer for Bimanual Dexterous Manipulation with a Plugin Tactile Adapter
- Precise Self-tuning of Spiking Patterns in Coupled Neuronal Oscillators
- Toward Practical Equilibrium Propagation: Brain-inspired Recurrent Neural Network with Feedback Regulation and Residual Connections
- A Neuromorphic VLSI System for Modeling the Neural Control of Axial Locomotion
- PointNetGPD: Detecting Grasp Configurations from Point Sets
- A Novel Simulation-Based Quality Metric for Evaluating Grasps on 3D Deformable Objects
- Minimal Work: A Grasp Quality Metric for Deformable Hollow Objects
- SoGraB: A Visual Method for Soft Grasping Benchmarking and Evaluation