PaperlayerView exact revision on arXiv

Source-linked AI summary

IORA: Irregular Object Immobilization under Restricted Actuation

Sang-Hun Han, Changyeon Won, JunGyu Lee, Min-Gyu Park, Hae-Gon Jeon

arXiv:2610.30969v1cs.RO

TL;DR

Grasp synthesis for disturbance-resistant dexterous hands often targets highly actuated hands, while coupled low-DoF hands may not realize the required contacts and support surfaces are often treated only as geometric constraints. This paper synthesizes support-aware grasps for coupled hands, reporting improved feasibility and robustness across two hands and executable real-robot grasps adaptable to downstream tasks.

  • Problem

    Disturbance-resistant grasp synthesis has focused on high-DoF hands, while coupled low-DoF hands may not realize required contact topologies and support surfaces are often treated only as geometric constraints.

    Source: §I · p. 1

  • Method

    The framework graph-matches hand and object regions subject to coupled-hand feasibility, then refines and physically validates grasps with support from the tabletop.

    Source: A. Graph-Based Contact Configuration · p. 3, §V · p. 8

  • Results

    Across the RH56 and Shadow Hand, accounting for hand embodiment and environmental support improves grasp feasibility and robustness, while real-robot grasps remain executable and adaptable to downstream tasks.

    Source: §V · p. 8

  • Takeaways & Limitations

    The synthesized grasps can remain executable on real robots and adapt to downstream tasks, within the evaluated setting.

    Source: §V · p. 8

    The current pipeline focuses on static tabletop support and known object geometry.

    Source: §V · p. 8

Abstract

from arXiv · show

Stable grasping under task-specific disturbances is important for robotic manipulation. However, most dexterous grasp synthesis methods are developed for highly actuated hands in free space, and low-DoF hands may not be able to realize the contact topologies required by these grasps. Meanwhile, support surfaces can provide useful contact forces, but are often treated only as geometric constraints. We present a support-aware grasp synthesis framework for object immobilization with low-DoF coupled hands. Given irregular object geometry, our method represents both hand and object local surface regions as graphs and searches for kinematically feasible matchings between them. We then refine the wrist pose and motor state for geometric contact feasibility and validates the refined grasps through physical rollouts. Starting from the validated grasp pose, an optional task-guided local search explores nearby grasp configurations by varying the wrist pose and motor state, and selects one according to the specified task condition. Experiments on diverse objects show that our method generates various grasp configurations that achieve high grasp success and strong disturbance resistance. Real-world experiments further demonstrate physically executable grasps which enable object immobilization in cooking scenario.

I. INTRODUCTION

Prior disturbance-resistant grasp synthesis focuses on high-DoF hands, while coupled low-DoF hands face contact-feasibility challenges. The framework uses tabletop support in synthesis and validation, then evaluates diverse grasps under disturbances and in real-world tasks.

I. INTRODUCTION

  • Prior disturbance-resistant grasp synthesis targets high-DoF hands, which are cost-prohibitive, while stable object pose is crucial for tasks such as cutting.The cited works focus on high-DoF dexterous hands with independently actuated joints.Source: §I · p. 1
  • Coupled low-DoF hands can produce unintended contacts because moving one region may move others and planned joint motions may exceed the hand’s limited degrees of freedom.These motions may push or displace the object during grasping.Source: §I · p. 1
  • The framework treats the support surface as an additional source of contact whose passive normal and frictional forces help a low-DoF hand resist external forces.The surface participates in grasp synthesis and physical validation, rather than serving only as a collision constraint.Source: §I · p. 1
  • The pipeline graph-matches hand and object regions, refines candidate contacts while preserving correspondences, validates grasps under disturbances, and optionally adapts validated grasps to task conditions.The optional local search varies wrist pose and motor state without repeating the global contact search.Source: §I · pp. 1–2
  • Experiments assess irregular objects under directional disturbances, compare against dexterous grasp methods, and demonstrate RH56 grasps supporting object immobilization during cutting and slicing.Real-object tests evaluate diverse grasp configurations and resistance to external disturbances.Source: §I · p. 2

A. Dexterous Grasp Synthesis

The paper contrasts prior grasp synthesis and environment-assisted methods with a framework that derives contact structures from the target hand’s coupled actuation. It represents grasp motion through wrist pose and motor states, and uses the tabletop as a load-bearing contact.

A. Dexterous Grasp Synthesis

  • Classical and learned dexterous-grasp methods often treat contact forces as friction-bounded variables, while this framework derives contact topologies from the target hand’s coupled actuation.It determines executability during synthesis without relying on a learned prior or seed grasp.Source: A. Dexterous Grasp Synthesis · p. 2

B. Underactuated and Coupled Hands

  • Prior underactuated-hand formulations constrain forces for a given contact set, whereas this method retains contact sets admitting a common coupled motor configuration.Reduced actuation restricts both reachable postures and realizable contact forces.Source: B. Underactuated and Coupled Hands · p. 2

C. Environment-Assisted Grasping

  • Unlike environment-assisted methods that use surfaces for acquisition, motion, or collision handling, this framework synthesizes static grasps with the support surface bearing load alongside the hand.The support remains load-bearing during synthesis and physical validation.Source: C. Environment-Assisted Grasping · p. 2

III. METHOD

  • The method represents a grasp by wrist pose T and motor state u, with fixed coupling mapping motor states to joint angles, and graph-based matching initiating contact synthesis.Refinement and disturbance validation follow the contact-matching stage; local task-guided search is optional.Source: §III · pp. 2–3, Figure 2 · p. 3

A. Graph-Based Contact Configuration

The method represents hand contact regions and object surface regions as graphs, then searches for feasible correspondences that yield diverse grasp initializations. Feasibility accounts for geometric relations and the coupled hand’s shared motor state.

A. Graph-Based Contact Configuration

  • Multiple feasible matchings define distinct contact configurations, which are converted into initial grasps by choosing a compatible motor state and aligning matched regions.The search explores different hand-region subsets and retains multiple solutions to preserve grasp diversity.Source: A. Graph-Based Contact Configuration · p. 3
  • Hand contact regions and local object surface regions form graphs whose edges encode distances and relative surface-normal angles.Hand nodes include fingertips, proximal finger surfaces, and palm surfaces; object regions are partitioned by spatial distance and normal consistency.Source: A. Graph-Based Contact Configuration · p. 3
  • A matching is feasible only when selected hand-region and object-region pairs have matching geometric relations under a common motor state.The method compares pairwise distances and relative normal angles, requiring one sampled motor state to satisfy all matched pairs.Source: A. Graph-Based Contact Configuration · p. 3
  • Object-region distance is represented as a range because contact locations can vary within each region.Pairwise relations include distance and normal angle.Source: A. Graph-Based Contact Configuration · p. 3

B. Region-Constrained Geometric Refinement

The method locally refines each graph-based grasp while keeping its hand–object region correspondence fixed. Its objectives improve contact feasibility while limiting penetration, unintended contact, and deviation from the initialization.

B. Region-Constrained Geometric Refinement

  • Refinement addresses penetration or separation that may remain in a kinematically consistent graph-based initialization.The grasp is represented by wrist pose and motor state, while the correspondence is held fixed.Source: B. Region-Constrained Geometric… · p. 3
  • Contact and penetration objectives improve matched-region contact while discouraging penetration into the object or support surface.The contact objective reduces separation and encourages matched surfaces to face each other.Source: B. Region-Constrained Geometric… · pp. 3–4
  • Clearance and regularization terms suppress unmatched contacts and keep wrist-pose and motor-state changes local.The fixed correspondence and coupled kinematics keep refinement within the hand’s realizable configuration space.Source: B. Region-Constrained Geometric… · p. 4
  • The correspondence remains unchanged, although exact contact locations may move within matched object regions, and geometric checks filter refined grasps before physical evaluation.Checks cover contact orientation, object penetration, tabletop clearance, and task constraints.Source: B. Region-Constrained Geometric… · p. 4

C. Physics-Based Validation and Selection

Refined candidates are validated by replaying grasp execution and holding in simulation under directional disturbances. Candidates passing the hold test are rechecked for geometric consistency and retained across contact structures.

C. Physics-Based Validation and Selection

  • Each refined candidate is replayed in simulation through approach, closure, and holding to evaluate stability under applied disturbances.During holding, hand contact forces and tabletop support reactions balance the object while disturbances are applied along six signed coordinate directions.Source: C. Physics-Based Validation and… · p. 4
  • A successful hold is assessed using the maximum object translation and rotation observed during the rollout.The corresponding limits are allowed translation and rotation thresholds.Source: C. Physics-Based Validation and… · p. 4
  • Candidates passing the hold test are rechecked for hand–object penetration and support-surface consistency, then retained across contact structures to preserve diversity.Source: C. Physics-Based Validation and… · p. 4

D. Optional Task-Guided Local Derivation

The optional local derivation adapts a physically validated grasp to task-specific requirements. It explores nearby wrist poses and motor states, enforces hard constraints, and ranks feasible candidates by task preferences.

D. Optional Task-Guided Local Derivation

  • Task-specific derivation starts from a physically validated grasp and explores nearby candidates by perturbing wrist pose and motor state.Perturbations include small translations, discrete yaw rotations, local motor-state scans, and contact-configuration changes.Source: D. Optional Task-Guided Local Derivation · p. 4
  • Hard task constraints define the feasible candidate set, while soft conditions specify preferences used to rank those candidates.Examples include keeping the hand outside a region, preserving an unobstructed load path, or requiring additional finger contacts; preferences include directional resistance, contact placement, and grasp style.Source: D. Optional Task-Guided Local Derivation · p. 4
  • The selected candidate is checked for geometric feasibility, adapting the grasp toward application requirements while reusing the validated grasp as its starting point.Source: D. Optional Task-Guided Local Derivation · p. 4

IV. EXPERIMENTAL RESULTS

The experiments evaluate grasp synthesis on shared simulation protocols and assess success, contact quality, penetration, and resistance to directional disturbances. They compare methods across hand embodiments and floating versus tabletop settings.

A. Experimental Setup

  • 200 DexGraspNet test-split objects are evaluated on coupled RH56 and Shadow hands under a shared MuJoCo protocol.Simulation uses a 1 ms step, 0.1 kg objects, friction coefficient 0.6, and default squeeze coefficient 0.5.Source: A. Experimental Setup · p. 4
  • GSR measures attempted grasps resisting disturbance, while OSR measures objects with at least one successful grasp.PD and SPD measure maximum hand–object and interlink penetration; CLN counts links within 2 mm, and lower CDC indicates more consistent finger contacts.Source: A. Experimental Setup · p. 4
  • Grasps are tested under forces along ±x, ±y, and +z, with comparisons against Dexonomy, DexGraspNet, and BODex.The protocol first tests 0.98 N for 2 s within displacement bounds, then increases force to 50 N at 8 N/s under tighter bounds; Fig. 3 reports weakest-direction resistance capped at 50 N.Source: A. Experimental Setup · pp. 4–5, Figure 3 · p. 5

B. Comparison with Existing Methods

On RH56, the method typically performs best across tabletop and floating settings, while the support surface strengthens resistance across larger disturbances. Results on Shadow Hand also show strong performance in both settings.

B. Comparison with Existing Methods

  • The method typically performs best on RH56 in tabletop and floating settings, with grasp success and disturbance resistance across the full force range.The comparison reports stronger resistance profiles with support across methods; Tab. I is identified as the quantitative comparison.Source: B. Comparison with Existing Methods · p. 5
  • Dexonomy's high GSR excludes difficult objects without valid grasps, although those objects count as failures for OSR.At matched grip forces below 20 N, median PD is similar to Dexonomy; higher mean PD reflects high-force cases above 20 N where Dexonomy often fails and is excluded.Source: B. Comparison with Existing Methods · pp. 5–6
  • On Shadow Hand, the method maintains high grasp success and strong disturbance resistance in both floating and tabletop settings.This embodiment test omits physics-based validation because GSR and OSR exceed 99% for nearly all methods when validation is used.Source: B. Comparison with Existing Methods · p. 6

C. Analysis and Ablations

The analyses examine the synthesis pipeline, support contacts, refinement objectives, local task-guided adaptation, and squeeze strength. Together, they report how these components relate to physical grasp performance and disturbance resistance.

C. Analysis and Ablations

  • Geometric contact synthesis establishes stable configurations but can produce substantial penetration; refinement reduces penetration, and physical validation restores grasp strength while filtering invalid grasps.Without physical validation, penetration-free grasps may lack sufficient contact force, while deeply penetrated grasps can remain successful.Source: C. Analysis and Ablations · p. 6
  • Removing tabletop contact substantially reduces force-balance feasibility and disturbance resistance while keeping the settled hand-object configuration unchanged.The comparison uses static SOCP analysis with and without tabletop contacts.Source: C. Analysis and Ablations · p. 6
  • The geometric refinement objective typically yields better physical grasp performance than calibrated SOCP and differentiable task-GWS alternatives.The alternatives optimize static wrench-based quality, which can be disconnected from contacts formed after closing and settling.Source: C. Analysis and Ablations · p. 6, Table V · p. 6
  • Direction-conditioned local derivation yields stronger resistance than the original grasps across each tested force direction.The local search adapts an already validated grasp to four horizontal target directions without repeating global contact search.Source: C. Analysis and Ablations · p. 6, Table VI · p. 6
  • Increasing squeeze strength consistently raises weakest-direction resistance while grasp success and settled penetration remain nearly unchanged.Median penetration remains constant across the squeeze range, indicating the increased resistance is associated with stronger contact loading rather than deeper penetration.Source: C. Analysis and Ablations · p. 6

D. Real-World Evaluation

Real-world tests evaluated grasp execution and resistance to pulley-applied loads, then demonstrated task-guided grasping during vegetable cutting. The experiments also show diverse physical grasp configurations across object shapes.

Real-world grasp execution

  • The RH56 physically executed diverse grasp configurations across object geometries, including fingertip, proximal-link, palm, and table-assisted contacts.The qualitative results show configurations adapting to different real-object shapes.Source: Figure 5 · p. 7, Figure 4 · p. 7

Real-world disturbance resistance

Downstream manipulation

  • For six real vegetables, task-guided local derivation selected support-assisted grasps that left the region needed for tool access clear during cutting.A knife or slicer was teleoperated while the RH56 held each object against the tabletop.Source: D. Real-World Evaluation · p. 7
  • As cutting changes an object's geometry, the framework can be reapplied to the updated shape to derive a new grasp for continued manipulation.Table VII evaluates squeeze strength in relation to resistance and settled penetration for fixed RH56 grasps.Source: Table VII · p. 7, D. Real-World Evaluation · p. 8

V. CONCLUSIONS

The conclusion presents support-assisted synthesis for low-DoF coupled hands, combining kinematically compatible contacts, geometric refinement, and physical validation. It identifies static tabletop support and known object geometry as current scope boundaries.

V. CONCLUSIONS

  • The framework searches for contact structures compatible with coupled-hand kinematics, refines them toward realizable contacts, and validates grasps through physical execution.It treats the support surface as an additional source of contact rather than only a geometric constraint.Source: §V · p. 8
  • Experiments on the RH56 and Shadow Hand show improved grasp feasibility and robustness when hand embodiment and environmental support are accounted for.Source: §V · p. 8
  • The current pipeline is limited to static tabletop support and known object geometry; broader environmental contacts and perception-driven geometries are proposed extensions.Source: §V · p. 8
Loading 2610.30969v1…