Source-linked AI summary
Relaxation-Aware Multimodal Sensing of Soft Gripper Driven by Structure-Perception-Learning
Yanzhe Wang, Hao Wu, Ziyi Zheng, Huixu Dong
TL;DR
Soft grippers must maintain interaction forces over extended holds despite compliance-related force instability and thermally coupled viscoelastic relaxation. This paper combines a variable-stiffness gripper, multimodal sensing, and physics-informed relaxation compensation. In a 280s force-controlled grasp-and-hold task, the proposed method achieved a mean absolute error of 0.066N and outperformed fixed-aperture and instantaneous-only baselines by 80% and 95%, respectively.
Problem
Soft grippers require stable long-duration grasp maintenance despite force decay from thermo-mechanically coupled viscoelastic relaxation during holding.
Method
The paper develops a variable-stiffness gripper with multimodal vision and infrared sensing, a temperature-coupled reduced viscoelastic model, and physics-informed force regulation.
Results
0.066N mean absolute error was achieved in a 280s force-controlled grasp-and-hold task, outperforming fixed-aperture and instantaneous-only baselines by 80% and 95%, respectively.
Takeaways & Limitations
The results support mechanism–AI co-design for stable, sustained soft grasping under coupled thermo-mechanical and viscoelastic dynamics.
Abstract
from arXiv · showhide
Achieving stable, sustained grasping with soft robotic hands remains a fundamental challenge. Compliance enables safe and adaptive contact, yet the intrinsic viscoelasticity of soft polymers leads to stress relaxation and a continuous decay of grasping force during holding. Inspired by human grasping, which combines phase-dependent stiffness regulation with continuous sensing and feedback, this paper presents an integrated structure--perception--learning framework. We develop a variable-stiffness soft gripper that uses onboard vision and infrared thermography to track deformation and the temperature field in real time, preserving continuous tracking of the interaction state. To mitigate relaxation-induced force decay, we propose a temperature-coupled viscoelastic force representation, together with a physics-informed learning model, to reconstruct the force trend and provide explicit compensation during holding. Experiments show that, in a 280s force-controlled grasp-and-hold task, the proposed method maintains the desired force with a mean absolute error of 0.066N, outperforming fixed-aperture and instantaneous-only baselines by 80% and 95%, respectively. Overall, the results support a mechanism--AI co-design view: mechanisms shape feasible interactions, while learning compensates remaining uncertainty in viscoelastic dynamics, together enabling stable, sustained grasping.
I. INTRODUCTION
Soft grippers combine robust, adaptive contact with force instability during prolonged holding because compliance and viscoelastic relaxation limit sustained force output. The paper addresses this gap through coordinated stiffness modulation, multimodal sensing, and physics-informed relaxation compensation.
- Compliance improves contact robustness but reduces load capacity and force stability during holding.
- Stress relaxation causes interaction forces to decay during sustained contact, with temperature governing both instantaneous stiffness and relaxation dynamics.
- Stable long-duration grasping requires compliant contact, tunable stiffness, continuous multimodal perception, and explicit compensation of time-dependent force evolution.
- The proposed bio-inspired finger combines passive compliance, thermally tunable stiffness, and switchable adhesion for multimode grasping across contact phases.
- Vision, infrared thermography, and model-based inference reconstruct interaction force from deformation and thermal observations.
- A temperature-coupled reduced viscoelastic representation and physics-constrained learning reconstruct force evolution and compensate relaxation during holding.
II. GRIPPER DESIGN
The gripper uses compliant skeletal fingers with thermally tunable stiffness, switchable adhesion, and base-mounted vision and thermal sensing. This architecture supports deformation-tolerant multimode grasping while preserving continuous observation across approach, contact, and holding.
- The system consists of two identical soft fingers integrating thermally tunable stiffness with base-mounted vision and thermal sensing.
- The open skeletal morphology enables large passive deformation and omnidirectional adaptation so fingers conform without precise kinematic alignment.
- Temperature-responsive SMP beams switch the finger between compliant conformal contact and stiff load-bearing holding without changing external geometry.
- Thermally activated surface pads provide switchable adhesion for planar or low-curvature surfaces alongside force-closure grasping.
- A base-mounted RGB camera tracks configuration through an ArUco marker, while an infrared camera monitors the temperature field during stiffness modulation.
III. PHYSICS-INFORMED FORCE PERCEPTION AND RELAXATION COMPENSATION
The force-perception framework models thermally dependent viscoelastic behavior during prolonged holding and uses multimodal observations to support force reconstruction and relaxation compensation.
- A generalized Maxwell–Wiechert formulation models the thermomechanical response of polymer-based soft fingers during prolonged holding.
- The model combines a temperature-dependent equilibrium spring with N parallel Maxwell branches containing temperature-dependent springs and dashpots.
- Under displacement-controlled contact, total interaction force combines the equilibrium elastic response with internal forces from the Maxwell branches.
- The holding phase is approximated as a step displacement input X(t) = X0 after contact establishment.
B. Reduced Temperature-Coupled Relaxation Representation
The paper reduces the temperature-dependent relaxation dynamics to a dominant effective mode that separates instantaneous force from subsequent relaxation loss. This representation is used within the learning and force-regulation pipeline for holding-phase compensation.
- The instantaneous contact force is defined at the onset of the holding phase.
- The second term in the rearranged force expression represents cumulative viscoelastic relaxation loss.
- Fast relaxation modes decay shortly after contact, while slow modes dominate force evolution over task-relevant holding horizons.
- The reduced model uses effective relaxation magnitude K(T) and time constant τ(T) to characterize temperature-dependent relaxation.
- The force-regulation pipeline combines instantaneous-force estimation, relaxation-parameter prediction, and closed-loop thermal and force control.
C. Physics-Aligned Learning and Force-Regulation Pipeline
The framework learns instantaneous contact force and temperature-dependent relaxation separately from multimodal observations, then uses the reconstructed force in closed-loop regulation. RGB and infrared sensing support online state tracking while a PID controller drives the gripper motor.
- Multimodal force perception: ICFN estimates instantaneous force from pose, squared marker translation, beam temperatures, temperature statistics, and spatial thermal gradients.These features encode deformation, thermal softening, and spatially non-uniform stiffness.
- Relaxation compensation: RCFN predicts relaxation magnitude and time constant from thermal and kinematic context during holding.Its inputs include current temperature, contact-onset pose, current pose, and pose velocity.
- Physics-aligned force representation: The reduced representation separates instantaneous contact response from time-dependent relaxation, making long-horizon force drift observable and compensable online.The approach learns reduced viscoelastic quantities rather than the full parameter set.
- Online sensing: RGB images update at approximately 30 Hz for pose and deformation tracking, whereas infrared images update at approximately 1 Hz for beam-level thermal feedback.The streams are converted into compact multimodal features for inference.
- Closed-loop regulation: The host PC fuses observations, reconstructs force through ICFN and RCFN, and sends it to a PID controller that commands the gripper motor through the MCU.A separate temperature-control loop regulates the thermal state of the fingers.
IV. EXPERIMENTS
The experiments evaluate multimodal force perception and long-duration grasp stability under thermally coupled viscoelastic effects. The evaluation focuses on the proposed variable-stiffness soft gripper and its relaxation-aware framework.
- Experimental scope: The experiments assess multimodal force perception accuracy and long-duration grasp stability under thermally coupled viscoelastic effects.The evaluation targets the performance of the proposed variable-stiffness soft gripper.
A. Temperature-Dependent Stiffness Characterization
Experiments characterize temperature-, position-, and load-dependent deformation before evaluating instantaneous force estimation and relaxation reconstruction. The results show thermomechanical softening, accurate force perception, and consistent relaxation modeling across temperatures and repeated contacts.
- Stiffness characterization: The finger’s temperature-dependent stiffness is characterized by measuring deformation under controlled normal indentation across 2–8N and 27–60◦C.The tested temperature range spans the glassy–rubbery transition.
- Thermomechanical response: Deformation increases monotonically with temperature across all loads, indicating pronounced thermomechanical softening.
- Thermomechanical response: A 1.6×–1.7× deformation increase from 27◦C to 60◦C corresponds to a 35%–45% reduction in k80 across force levels.At 8N, global deformation increases from 361mm2 to 626mm2, or +73%.
- Modeling basis: The coupled spatial–thermal deformation patterns provide a physical basis for modeling instantaneous contact force and temperature-dependent viscoelastic relaxation.The patterns vary with contact position, temperature, and load.
- Evaluation metrics: The evaluation reports pointwise force errors for ICFN and curve-level NRMSE and AUCrel metrics for RCFN.AUCrel quantifies relative cumulative drift, while Amp = F0 − Fend characterizes relaxation strength.
- Force perception: ICFN achieves an MAE of 0.651 N, with RMSE of 0.886 N and R2 = 0.848 over the 0–10 N force range.Median absolute errors remain bounded from weak to high-load contacts.
- Relaxation reconstruction: RCFN attains mean/median NRMSE of 0.160/0.118, mean/median AUCrel of 0.065/0.049, and mean curve-level correlation of 0.990.Representative reconstructions capture both relaxation magnitude and temporal evolution.
- Repeatability: Across repeated contacts, relaxation reconstruction remains consistent at approximately 40◦C and 50◦C, while correlation remains high at approximately 60◦C despite increased spread.At approximately 60◦C, Corr = 0.991 ± 0.013.
C. Holding Force Measurement and Relaxation-Aware Force Control
The proposed ICFN+RCFN method models relaxation-induced force evolution and compensates it during holding, maintaining stable regulation around a 5 N target over approximately 280 s.
- The fixed-aperture baseline decayed steadily, while instantaneous-only estimation increasingly overestimated true force as relaxation progressed.
- 5 N target regulation remained stable throughout holding with the proposed relaxation-aware method.
- 0.066 N MAE represented 1.32% of the 5 N target, compared with 0.333 N for fixed-aperture and 1.327 N for instantaneous-only baselines.The proposed method reduced MAE by approximately 80.2% and 95.0%, respectively.
- Explicitly modeling and compensating temperature-dependent viscoelastic relaxation improved force perception and long-duration grasp stability.
D. Real-Time Force Perception for Stiffness-dependent Stable Grasping
Variable stiffness changes force-response dynamics and feasible interaction forces, while real-time perception supports stable closed-loop grasping across stiffness regimes.
- Instantaneous force estimates remained stable across high-, medium-, and low-stiffness states despite stiffness-dependent deformation patterns.Lower temperatures produced higher stiffness and larger, steeper contact-force transients.
- Softer configurations enhanced sensitivity to weak contacts, whereas stiffer configurations favored larger force generation and maintenance.Feather and brush contacts produced clearly detectable signals.
- 5 N hammer grasping at high stiffness, 6 N orange grasping at medium stiffness, and 0.3 N balloon grasping at low stiffness demonstrated task-specific regulation.
- Estimated force stayed tightly regulated around desired targets through approach, closure, holding, and lifting phases.
- Stiffness modulation constrained feasible force ranges and interaction dynamics, while force perception supplied feedback for force-closure grasping.
E. Integrated Demonstration of Variable Stiffness and Multimodal Perception
An integrated manipulation sequence combines thermal identification, visual proximity sensing, variable stiffness, and force control to handle objects requiring different interaction conditions.
- The task combined non-contact cup identification, compliant grasping and pouring, and stiff glass-cup grasping for human handover.The stages imposed distinct physical requirements that a fixed stiffness could not satisfy.
- Infrared sensing discriminated room-temperature and hot-water cups before visual proximity-guided approach.Finger-mounted RGB cameras provided visual proximity cues.
- 1 N compliant grasping enabled paper-cup lifting and pouring while reducing contact stress and accommodating misalignment.
- Lowering temperature increased stiffness for 6 N glass-cup grasping and handover, improving load capacity and disturbance resistance.
- The full force profile showed coordinated perception, approach, compliant grasping, stiffness transition, and handover.Variable stiffness switched interaction regimes while multimodal perception and force control maintained adaptive manipulation.
V. CONCLUSION
The paper integrates a variable-stiffness soft gripper, multimodal sensing, and physics-informed relaxation compensation for stable long-duration grasping under thermo-mechanical viscoelasticity.
- The bio-inspired finger combines passive compliance, thermally tunable stiffness, and switchable adhesion for multimode grasping.Base-mounted vision and thermal sensing preserve interaction observability under deformation.
- A temperature-coupled reduced viscoelastic model separates instantaneous force from relaxation-induced decay.
- Physics-informed learning reconstructs relaxation-driven force evolution and supplies explicit compensation for long-horizon closed-loop stabilization without embedded force sensors.
- Experiments demonstrated robust contact acquisition, stiffness modulation, force perception, and improved holding stability over fixed-aperture and instantaneous-only baselines.