PaperlayerView exact revision on arXiv

Source-linked AI summary

A Support-Enhanced Granular-Jamming Gripper for RL-based Grasping with Continuum Manipulators

Danyu Liu, Tianlin Zhang, Wei Chen, Wei Tang, Kecheng Qin, Zhongyu Li

arXiv:2609.29093v1cs.RO

TL;DR

Residual position and orientation errors can leave continuum manipulators without the pose needed for rigid grasping. This paper combines a support-enhanced granular-jamming gripper with learned reaching and demonstrates autonomous physical grasping and transfer, while identifying geometry-dependent grasp boundaries.

  • Problem

    Residual tip position, orientation, and motion errors can cause collision or prevent stable contact when rigid grasping requires a suitable relative pose and geometric compatibility.

    Source: §I · p. 1

  • Method

    The paper pairs a compliant, support-rod granular-jamming gripper with visual feedback, pneumatic task control, and a randomized-simulation PPO reaching policy.

    Source: §I · pp. 1–2, §VIII · p. 7

  • Results

    10/10 lift-and-hold trials succeeded with the support rod versus 6/10 without it; the integrated system demonstrated autonomous reach, grasp, lift, transfer, and release.

    Source: §VIII · p. 7

  • Takeaways & Limitations

    The results establish that mechanically adaptive grasp formation can operate within an autonomous continuum-manipulation system.

    Source: §VIII · p. 7

    Graspability remains geometry-dependent, evaluation covers limited approach conditions and object geometries, and the internal support geometry is not fully optimized.

    Source: §VIII · p. 7

Abstract

from arXiv · show

Continuum manipulators provide dexterous motion in confined spaces, but structural compliance, hysteresis, and load-dependent deformation leave residual position and orientation errors that can undermine reliable contact with rigid grippers. To address this limitation, this paper presents a lightweight support-enhanced granular-jamming gripper tailored to a continuum manipulator. The gripper maintains compliance before jamming while establishing a direct load path to the continuum manipulator tip after jamming. To improve its grasping performance, we systematically designed membrane materials, particles, filling ratios, and the internal support structure, and further identify geometry-dependent grasp boundaries with respect to contact offset and object shape. Building on these results, we construct a physical manipulation system integrating the continuum manipulator, granular-jamming gripper, visual feedback, tendon actuation, and pneumatic control. We then train a reinforcement-learning-based reaching controller in a randomized simulation and deploy it on the physical system, demonstrating how positioning control and contact level mechanical adaptation can complement each other in a modular grasp-and-release task. By introducing an adaptive structure that relaxes the need for highly accurate modeling and positioning control, this work explores a design paradigm that integrates physical and embodied intelligence.

I. INTRODUCTION

Residual position and orientation errors make rigid grasping difficult for continuum manipulators, while adaptive contact must remain lightweight and compatible with uncertain physical reaching. The paper introduces a support-enhanced granular-jamming gripper and an integrated system that pairs learned reaching with mechanically formed grasping.

I. INTRODUCTION

  • Residual tip errors can cause collision, loss of contact, or failure to establish a stable constraint when a rigid gripper requires a suitable relative pose and geometric compatibility.Visual feedback can reduce positioning error, but finite position, orientation, and motion residuals may remain.Source: §I · p. 1, Figure 1 · p. 1
  • The reaching policy must tolerate physical differences in posture, stiffness, actuation timing, and target location while bringing the gripper into a contactable region.It is not assumed to eliminate every contact-level error.Source: §I · p. 1
  • The lightweight gripper conforms during contact, uses an internal support rod to promote enclosure and transmit compression, and provides a direct load path after jamming.Its geometry-dependent grasp boundaries are characterized across contact offsets and object shapes.Source: §I · pp. 1–2
  • 10/10 lift-and-hold trials succeeded with the support rod versus 6/10 without it; terminal RMSE was 4.23 cm over 16 physical rollouts versus 2.51 cm in simulation.The work also demonstrates the complete physical reach–grasp–lift–transfer–release sequence.Source: §I · p. 2

II. RELATED WORK

  • The related-work discussion organizes approaches to continuum-manipulation uncertainty around continuum modeling and control, mechanically adaptive grippers, and learning-based sim-to-real transfer.It focuses on how these approaches reduce or accommodate uncertainty during reaching and grasping.Source: §II · p. 2

A. Continuum-Manipulator Modeling and Control

The related work describes complementary ways to manage uncertainty in continuum manipulation: model-based and feedback control, mechanically adaptive grasping, and learning-based sim-to-real transfer.

A. Continuum-Manipulator Modeling and Control

  • Analytical continuum models support analysis and control, but their accuracy depends on mechanical parameters and loading conditions that may vary during physical operation.Feedback from measured states reduces this dependence, while the proposed approach adds end-effector adaptability to enlarge successful contact configurations.Source: A. Continuum-Manipulator Modeling and… · p. 2

B. Mechanically Adaptive Grippers

  • Adaptive grippers capture objects through deformation, wrapping, or enclosure, but their deformation paths, dimensions, and actuation needs can complicate integration with slender continuum manipulators.Granular jamming retains passive conformity and pneumatic locking; the internal support adds lateral enclosure and a more direct load path after jamming.Source: B. Mechanically Adaptive Grippers · p. 2

C. Learning-Based Control and Sim-to-Real Transfer

  • The reaching formulation combines randomized posture, target position, stiffness, and actuation delay with four-frame observations and absolute tendon-length commands for zero-shot deployment.A policy trained in simulation may otherwise depend on dynamics that do not hold on physical hardware.Source: C. Learning-Based Control and… · p. 2

III. SYSTEM OVERVIEW

The system separates learned tendon-based reaching from pneumatic task control and mechanically adaptive grasp formation. Its task sequence uses continued contact to form an enclosure before jamming, lifting, transfer, and release.

III. SYSTEM OVERVIEW

  • Visual feedback guides tendon actuation toward the object, while continued contact reshapes the unjammed gripper before vacuum locks the configuration for lifting and transfer.The reaching policy need only approach a finite contactable region rather than a unique exact grasp pose.Source: §III · p. 3, Figure 2 · p. 3

A. System Architecture

  • The architecture combines a tendon-driven continuum manipulator, distal granular-jamming gripper, tendon and pneumatic actuators, an RGB camera, and a control computer.The learned policy controls reaching, the task layer coordinates pneumatic switching, and the gripper mechanically forms and retains contact.Source: A. System Architecture · p. 3

B. Autonomous Task Sequence

  • After the reaching condition is met, continued approach develops contact and enclosure before vacuum is applied; the manipulator then lifts, transfers, and releases the object at a designated region.The destination is a region rather than a unique pose.Source: B. Autonomous Task Sequence · p. 3

IV. THE DESIGN OF GRANULAR-JAMMING GRIPPER

The gripper is designed to remain compliant during contact while transmitting force through an internal support and retaining the grasp after jamming. Its operating sequence uses contact-driven conformity followed by evacuation.

A. Structure and Principle

  • The support-enhanced structure combines a conformable membrane and granular filler with an internal support that promotes lateral enclosure, transmits compressive force, and provides a direct load path after jamming.These functions suit the gripper to a bending continuum tip.Source: Figure 4 · p. 3, A. Structure and Principle · pp. 3–4
  • Before evacuation, mobile particles allow contact to reshape the flexible membrane, which confines the granular filler.The design evaluation considers the membrane–particle combination, filling ratio, and support configuration.Source: A. Structure and Principle · p. 3, §IV · p. 3
  • Continued compression redistributes particles and deepens enclosure before evacuation increases interparticle constraints and locks the adapted configuration.Load bearing does not begin at first contact; the unjammed state provides conformity, while the jammed state provides retention during lifting and transfer.Source: A. Structure and Principle · p. 4, Figure 5 · p. 4

B. Materials Configuration

The configuration study selects a natural-rubber membrane, TPR spheres, a 50% nominal filling ratio, and an internal support rod. The choices balance measured grasp success with distal load and enclosure mechanics.

Configuration Study

  • Each configuration condition used ten trials with a common object, evacuation condition, operating procedure, and success criterion.A trial counted as successful when the object was lifted clear of the table and retained through the lifting phase.Source: B. Materials Configuration · p. 4

Membrane and Particle Selection

  • With a natural-rubber membrane, PS particles and TPR spheres each achieved 100% grasp success; TPR was selected for lower total gripper mass.The lower mass reduces distal load on the compliant continuum manipulator.Source: B. Materials Configuration · p. 4

Filling-Ratio Selection

  • Success rates with natural rubber and TPR spheres were 90%, 100%, and 70% at filling ratios of 1/3, 1/2, and 4/5, respectively.Low filling leaves more deformation space but fewer load-bearing particles; excessive filling restricts conformity and increases distal load, so the final nominal filling ratio is 50%.Source: B. Materials Configuration · p. 4

Support-Rod Evaluation

  • The supported gripper succeeded in 10/10 trials, versus 6/10 without the rod under otherwise matched conditions.The support rod maintains useful contact geometry, transmits compressive force to promote enclosure, and provides a direct load path after jamming; the experiment evaluates these coupled functions together.Source: B. Materials Configuration · p. 4

V. THE CHARACTERIZATION OF THE GRIPPER

The characterization tests grasping with contact offsets while separating gripper mechanics from learned reaching control. Results show that successful contact depends on geometry, support-rod posture, and the surrounding table constraint rather than on a uniform tolerance region.

A. Grasping Under Contact-Position Offset

  • Grasping was tested at discrete lateral and longitudinal offsets around O, the table projection of the gripper sphere’s geometric center, which was not assumed to be optimal.Continuum motion was manually controlled or fixed to isolate gripper mechanics from the learned reaching controller.Source: A. Grasping Under Contact-Position… · p. 4, §V · p. 4
  • The high-success region shifts away from O because the 30°–40° downward-inclined rod and table help form a joint enclosure on the object’s distal side.Continued approach force helps the membrane and particles develop a deeper enclosure in favorable configurations.Source: A. Grasping Under Contact-Position… · pp. 4–5

V. THE CHARACTERIZATION OF THE GRIPPER

  • The selected configuration succeeded in 10/10 trials with the support rod and 6/10 without it, with ten trials per condition.The comparison evaluates the selected configuration under the tested condition.Source: Figure 6 · p. 5

A. Grasping Under Contact-Position Offset

  • Figure 7’s continuous field interpolates measurements at discrete positions for visualization; it does not represent measurements at every location.The membrane, support rod, object, and table jointly form a stable enclosure under favorable configurations.Source: A. Grasping Under Contact-Position… · p. 4, Figure 7 · p. 5
  • The highest measured offset-condition success rate was 100%; success was 80% near O, reached 90% at selected lateral offsets, and fell to about 50% or below without stable joint constraint.The measured region depends on support-rod posture and environmental constraint; other approach orientations may shift it and remain to be tested.Source: A. Grasping Under Contact-Position… · p. 5

B. Grasping Objects of Different Shapes

Grasp success varies across object geometries, with the experiments linking enclosure performance to available contact area and shape. The following section frames reaching control as bringing the gripper into a contactable state, not controlling grasp formation or pneumatic switching.

B. Grasping Objects of Different Shapes

  • The sphere was grasped in 10/10 trials (100%), and the hexagonal prism in 9/10 trials (90%).Each shape was tested in ten trials under the common protocol.Source: B. Grasping Objects of Different Shapes · p. 5
  • The cylinder and cube each succeeded in 8/10 trials (80%), while the triangular pyramid succeeded in 1/10 trials (10%).The successful pyramid trial still formed a load-bearing enclosure.Source: B. Grasping Objects of Different Shapes · p. 5
  • Continuous surfaces support distributed enclosure, while sharp pyramid edges and a small stable contact region promote folding, slip, or rotation before enclosure develops.Cubes and hexagonal prisms still provide sufficient contact area and geometric constraint for load-bearing configurations.Source: B. Grasping Objects of Different Shapes · p. 5
  • The contact-offset and shape experiments show that compliance enlarges, but does not make unrestricted, the set of configurations where a load-bearing enclosure can form.The enclosure depends jointly on the membrane, particles, support rod, and environment.Source: B. Grasping Objects of Different Shapes · p. 5

VI. CONTINUUM MODELING AND REINFORCEMENT-LEARNING-BASED REACHING

  • The reaching controller brings the gripper into a contact-suitable state rather than controlling grasp formation or pneumatic switching.The described approach formulates reaching with a nominal PCC model, history-based observation, absolute tendon-length actions, targeted domain simulation, and PPO training for zero-shot physical deployment.Source: §VI · p. 5

A. Continuum Modeling

The piecewise-constant-curvature model represents each continuum segment with constant curvature and bending-plane angle. It supplies a low-dimensional geometric basis for policy training, not an exact model of physical dynamics.

A. Continuum Modeling

  • PCC modeling assumes each segment’s curvature and bending-plane angle remain constant along its arc length.The model describes the centerline position for a segment of length L.Source: A. Continuum Modeling · p. 5
  • Across ten trials per shape, success rates were 80% for cylinders, 80% for cubes, 100% for spheres, 90% for hexagonal prisms, and 10% for triangular pyramids.The results identify sharp-edged objects with limited stable contact as a design limitation.Source: Figure 8 · p. 6
  • The PCC model provides a low-dimensional geometric basis for policy training but does not fully reproduce physical continuum dynamics.The model omits effects including tendon friction, hysteresis, material variation, actuation delay, external loading, and coupled deformation.Source: A. Continuum Modeling · p. 6

B. Policy Observation and Action

The reaching policy uses a shared base-fixed task frame and four-frame observation history to issue absolute tendon-length commands. Its reward trains approach, slowing near the target, and maintaining a contactable state.

B. Policy Observation and Action

  • A base-fixed task frame centered at the manipulator base gives camera-detected gripper and target positions a consistent representation across simulation and deployment.All position quantities are expressed in this frame.Source: B. Policy Observation and Action · p. 6
  • Each observation combines gripper position, relative target displacement, and tendon state; stacking four observations supplies short-term motion context.The history is associated with hysteresis, stiffness variation, and actuation delay.Source: B. Policy Observation and Action · p. 6
  • The policy outputs four bounded absolute tendon-length targets, while visual feedback and observation history help it respond to posture deviations caused by friction, hysteresis, stiffness variation, and loading.Absolute targets avoid incremental-command drift under delayed feedback, though identical targets can produce different postures.Source: B. Policy Observation and Action · p. 6

C. Reward Design

  • The reward combines distance-reduction, target-neighborhood, slow-motion, and hold terms to train approach, slowing, and sustained presence near the target.The hold term prevents a transient pass through the target region from counting as successful reaching.Source: C. Reward Design · p. 6

D. domain simulation

A randomized PCC-based simulation trains a history-based PPO reaching policy whose observation and action meanings carry over to the physical system without fine-tuning. The integrated system completes autonomous grasping and release while keeping learned reaching separate from mechanical contact adaptation.

D. domain simulation

E. PPO Training and Deployment

  • Physical deployment replaces simulated positions with camera measurements in the same base-fixed frame while retaining the state structure and absolute tendon-length action, with no physical-system fine-tuning.Grasp formation and pneumatic switching remain outside the learned controller.Source: E. PPO Training and Deployment · p. 7, Figure 9 · p. 7

VII. AUTONOMOUS GRASP-AND-RELEASE EXPERIMENTS

  • The integrated physical system completed the reach–grasp–lift–transfer–release sequence using visual state estimation, learned tendon control, passive grasp formation, and task-level pneumatic switching.The gripper remained unjammed during approach, conformed during contact, retained the object after evacuation, and released it over the destination.Source: §VII · p. 7

VIII. CONCLUSION AND FUTURE WORK

The support-enhanced gripper conforms before evacuation and uses its support rod and jammed particles for retention and load transfer afterward. Its grasp performance varies with geometry, while integration with a zero-shot reaching policy enables an autonomous grasp-and-release sequence within the tested scope.

VIII. CONCLUSION AND FUTURE WORK

  • Before evacuation, the flexible membrane and mobile particles conform to the object while the support rod promotes lateral enclosure and transmits compressive force; after jamming, the structure provides retention and load transfer.The support rod configuration achieved 10/10 lift-and-hold trials, compared with 6/10 without the rod.Source: §VIII · p. 7
  • Grasp performance is geometry-dependent: the gripper enlarges a set of successful contact configurations rather than providing isotropic position tolerance, and triangular pyramids remain difficult.The work systematically evaluated membrane material, particle material, filling ratio, and support structure.Source: §VIII · p. 7
  • The integrated system demonstrated the complete autonomous reach–grasp–lift–transfer–release sequence using randomized-simulation reaching without physical fine-tuning.The policy does not need to model or control granular contact deformation.Source: §VIII · p. 7, Figure 10 · p. 8
  • The prototype’s performance has been evaluated under only a limited set of approach conditions and object geometries, and its support structure remains unoptimized.Future work includes broader three-dimensional position and orientation variation and contact or pressure sensing.Source: §VIII · pp. 7–8

Related papers

Loading 2609.29093v1…