How Much Does a Tendon Tell You? Information-Optimal Coupling for Underactuated Hands

Cheap robot hands get cheap by sharing motors. They tie two finger joints to one tendon, fuse the sideways spread of two fingers, or let one motor close the whole hand. Designers pick these couplings by copying anatomy or by trial and error. They then judge the result by whether the hand can still reach a set of grasp poses or hold an object. That judgment leaves something out. A coupled hand that closes passively around an object takes on a shape that depends on the object, so the coupling itself does part of the sensing. Some couplings make that shape informative about what is being touched. Others squash different objects into the same posture and leave the sensors with nothing to report.

We propose treating the mechanism, its sensors, and the first stage of sensory processing as a single noisy communication channel. We would choose the coupling, under a fixed motor budget, to maximize the mutual information between the contact situation and what the hand's encoder reports. Mutual information is the number of bits the output carries about the input. This is the "infomax" principle that Linsker proposed for tuning synaptic weights in layered perceptual networks. Here we extend it from weights to tendons and joints. The claim we want to test is practical. An information score can be computed from short passive closures, without running full grasping policies. We expect it to rank hand designs the way closed-loop tactile manipulation success ranks them, and to find good designs with far fewer expensive evaluations than optimizing grasp success directly.

Core Claim

Information-theoretic plasticity work has always held the sensors fixed and optimized what comes after them. The clearest example models second-order touch neurons in the cuneate nucleus, a relay station in the brainstem (Optimal context separation of spiking haptic signals by second-order somatosensory neurons). It applies an information-maximizing spike-timing rule to about 1,700 recorded human fingertip afferents across 81 contact conditions that vary curvature, force magnitude, force direction, and angle. The rule cuts spike-timing jitter from about 12 ms to 0.6 ms and reaches full discrimination within 35 ms. Robot-hand co-design does the opposite. It changes the mechanics but scores them only by task reward. Emergent Hand Morphology and Control from Optimizing Robust Grasps of Diverse Objects co-optimizes finger count, link lengths, and torques with latent-space Bayesian optimization (LABO) on grasp success, and finds three fingers to be the best tradeoff.

Our proposal sits between these two. We optimize the coupling parameters $\theta$ and an encoder $E_\phi$ jointly:

$$\max_{\theta,\phi}\; I\big(X;\, E_\phi(S_\theta(X,u))\big) \quad \text{subject to} \quad \mathrm{DoA}(\theta)\le k$$

The symbols are as follows. $X$ is the contact situation: object shape class, contact location, and force direction, structured like the 81-condition design. $u$ is the low-dimensional closing command. $S_\theta$ is the joint-angle, tendon-tension, and fingertip-pressure reading produced by a hand with coupling $\theta$. $k$ is the number of motors allowed. We estimate $I$ with a contrastive lower bound (InfoNCE). This sidesteps the problem the haptic paper solved with spike-train distance metrics, namely that mutual information is hard to compute for continuous high-dimensional signals.

The design question is concrete. When two joints share a tendon, information about the object moves out of the motor commands, which the controller already knows, and into the passive joint configuration and tendon tension. The best coupling is the one whose passive conformation separates contact situations most cleanly. We predict that information plotted against motor count will show a knee: bits gained per added motor will drop sharply beyond a small number of motors. This mirrors the population-size knee near 50 neurons in the cuneate model, and it would be an information-theoretic account of the diminishing returns that the DoF literature documents mechanically.

Research Program

The parameterized design family comes straight from real hands. It includes a ratio for coupling the middle and end finger joints, as in RUKA, which drives 15 joints with 11 motors, and in A Wearable Robotic Hand for Hand-over-Hand Imitation Learning. It includes optional fusion of the fourth and fifth finger spread, as in the SABD hand. It includes single-tendon routing across multiple joints with adjustable pulley radii and spring stiffnesses, after the cable-driven 3D-printed hand. Finally, it includes a two-motor shared-tendon option in which the motor difference changes how tension is distributed along the tendon through friction, as in the Pisa/IIT SoftHand+. The total motor count runs from 1 to 11.

For each design, we run passive closures in BulletPhysics on the 48 EGAD objects at sampled poses and contact offsets. We inject realistic sensor noise and tendon hysteresis, because LABO's purely rigid-body evaluation is a known weakness. We then train a small encoder to maximize the InfoNCE bound. The outer loop searches over $\theta$ with the same latent Bayesian optimizer LABO uses, with information replacing grasp success as the score. This keeps the optimizer fixed and isolates the effect of the objective.

As a second line, we compare the contrastive encoder with an information-maximizing spiking front end in the style of Bohte et al.. That rule reduces output uncertainty given the input, which suppresses jitter. The comparison asks whether sparse event coding adds anything for a robot. We treat it as an optional comparison, not a core commitment.

Baselines

The naive reference is the anthropomorphic or uniform choice: equal coupling ratios, no fusion, one tendon per finger, evaluated at each motor count. Uniform random sampling of designs gives the floor; the LABO paper reports 7% overall grasp success for it on EGAD.

The strongest existing method is LABO optimizing grasp success directly, which reports 50% overall grasp success on EGAD, against 37% for Bayesian optimization on raw parameters and 26% for CMA-ES (Emergent Hand Morphology and Control). It is the strongest competitor because it already solves co-design in this simulator on this object set. To keep the comparison fair, we rerun LABO inside our coupling family with the same simulator noise, the same objects, and the same total number of physics steps. Our passive closures are cheaper per evaluation than full grasp trials, so we match on simulation compute rather than on evaluation count. We also report results matched on evaluation count. We beat LABO if, at equal compute, designs chosen by information match or exceed its grasp success when a learned tactile policy controls the hand. A stronger win would be reaching LABO's success level with substantially less compute.

The key ablation is variance-only design selection. We score designs by the total variance or autoencoder reconstruction quality of the sensor readings rather than by information about $X$. This tells us whether the gain comes from task-relevant information or merely from "more signal." A second ablation removes the motor-count constraint, which tests whether a fully actuated design at the same sensor budget closes the gap.

Why not just…?

Why not just optimize grasp success, as LABO does? This is the main threat. Success is the quantity we ultimately care about, and LABO already improves on raw Bayesian optimization from 37% to 50%. Our argument is that success is sparse and requires a trained controller for every candidate design. Information is a dense score available from passive closures, and it measures what a closed-loop tactile controller could exploit, which open-loop success does not. Experiment one settles the question directly. If information-based rankings do not predict tactile-policy success better than grasp-success rankings at equal compute, the proposal fails.

Why not just add motors and sensors and let software handle coordination? RUKA reports a 6.0 kg payload for its coupled design against 4.0 kg for LEAP and 3.6 kg for Allegro. The weight cascades documented in that line of work make full actuation costly at the wrist. Our motor-count sweep tests the claim in information terms. If bits keep rising linearly with motors, with no knee, the argument for coupling weakens, and we would report that.

Why not just compress the touch signals with an autoencoder or principal components? Reconstruction keeps whatever varies most, including tendon-friction noise. Information about $X$ keeps only what distinguishes contact situations. The variance-only ablation measures this difference. If it matches the information score, then the infomax framing adds nothing and we would say so.

Proposed Experiments

The first decisive check is rank agreement. For about 60 designs spread across motor counts, we train a tactile-plus-proprioception grasping policy on EGAD. We then measure how well rank correlation with that policy's success is predicted by the information score, by LABO-style open-loop success, and by sensor variance, each computed at equal compute.

The second check is optimization. We compare LABO with an information objective against LABO with a success objective, reporting final closed-loop success and compute needed to reach LABO's 50% level.

The third check is the knee. We plot best-achievable information against motor count from 1 to 11, and test whether the knee falls near the three-finger, few-motor regime that LABO found by reward alone. Agreement would suggest that reward-optimal morphology is largely information-optimal morphology.

The fourth check isolates the encoder half on real data. On DECO-50, we replace DECO's LoRA tactile adapter input with our infomax-pretrained compressed tactile code. The reference results are 82.50% average success for the tactile-adapter model and 72.25% for vision only (DECO). Matching 82.50% with a much smaller tactile code would show that the information front end transfers. Falling to the vision-only level would show that it discards what policies need.

Finally, we 3D-print the best and worst information-ranked finger couplings with low-cost fingertip pressure pads. We measure classification accuracy on contact situations from the real hand, to check that the simulated ranking holds up against real tendon friction.