PaperlayerView exact revision on arXiv

Source-linked AI summary

Force Policy: Learning Hybrid Force-Position Control Policy under Interaction Frame for Contact-Rich Manipulation

Hongjie Fang, Shirun Tang, Mingyu Mei, Haoxiang Qin, Zihao He, Jingjing Chen, Ying Feng, Chenxi Wang, Wanxi Liu, Zaixing He, Cewu Lu, Shiquan Wang

arXiv:2602.22088v2cs.RO

TL;DR

Contact-rich manipulation needs global visual guidance and stable local force control, yet prior policies often entangle these roles or assume the interaction structure. The paper introduces a physically grounded interaction frame recovered from demonstrations and Force Policy, which combines global vision with high-frequency hybrid force-position control. Across diverse tasks, it reports improved robustness, force regulation, and generalization to novel objects, while its interaction-frame theory excludes highly non-smooth and non-conservative contact transitions.

  • Problem

    Existing policies often entangle global perception with local contact refinement, while control-centric methods commonly assume interaction structure or learn controller parameters instead of the structure itself.

  • Method

    Force Policy uses a physically grounded interaction frame recovered from demonstrations, with a global vision policy for task progress and a high-frequency force policy for hybrid force-position control during contact.

  • Results

    Force Policy consistently outperforms vision-only and force-aware baselines across diverse contact-rich tasks, with high success on unseen objects where baselines frequently suffer catastrophic failure.

  • Takeaways & Limitations

    The global-local design combines task-level generalization with stable contact execution under sensor noise, fast contact transients, and varied object properties.

  • Takeaways & Limitations

    The interaction-frame theory covers stable, locally conservative frictional and compliant contacts but excludes highly non-smooth or non-conservative transitions such as fracture.

Abstract

from arXiv · show

Contact-rich manipulation demands human-like integration of perception and force feedback: vision should guide task progress, while high-frequency interaction control must stabilize contact under uncertainty. Existing learning-based policies often entangle these roles in a monolithic network, trading off global generalization against stable local refinement, while control-centric approaches typically assume a known task structure or learn only controller parameters rather than the structure itself. In this paper, we formalize a physically grounded interaction frame, an instantaneous local basis that decouples force regulation from motion execution, and propose a method to recover it from demonstrations. Based on this, we address both issues by proposing Force Policy, a global-local vision-force policy in which a global policy guides free-space actions using vision, and upon contact, a high-frequency local policy with force feedback estimates the interaction frame and executes hybrid force-position control for stable interaction. Real-world experiments across diverse contact-rich tasks show consistent gains over strong baselines, with more robust contact establishment, more accurate force regulation, and reliable generalization to novel objects with varied geometries and physical properties, ultimately improving both contact stability and execution quality. Project page: https://force-policy.github.io/

I. INTRODUCTION

Contact-rich manipulation requires coordinated global visual guidance and local force feedback, but existing approaches often entangle these roles or assume the interaction structure. Force Policy addresses this gap with a global-local vision-force design centered on an explicit interaction structure.

  • Motivation: Contact-rich tasks require simultaneous motion control and force modulation under uncertain geometry, friction, and compliance.These demands arise in tasks including assembly, peeling, polishing, and other physical interactions.
  • Motivation: Human-like organization separates global vision for task progress from high-frequency force feedback for stable local contact.The global component should generalize across task variations, while the local component refines execution and maintains contact.
  • Prior limitations: Monolithic learning policies jointly learn perception, planning, and contact refinement, making it difficult to preserve a dedicated high-frequency interaction loop while improving global generalization.Some methods append force signals or use them as auxiliary supervision, while slow-fast designs can tightly couple the fast component to the slow policy.
  • Prior limitations: Control-centric methods commonly rely on task-specific models and manual tuning, while learning-based variants usually predict controller parameters without representing the interaction structure itself.Prior work includes hybrid force-position and impedance control, learned admittance gains, and predicted contact forces.
  • Approach: Force Policy makes the missing structure explicit through a global-local vision-force policy with phase-wise authority.A global vision policy guides task progress, while a high-frequency force module estimates the interaction frame and executes hybrid force-position control during contact.

C. Control Structure Discovery from Demonstrations

The paper generalizes task-frame decomposition into a physically grounded interaction frame defined by environmental response and task intent. It recovers this frame from interaction signals and uses orthogonalization to handle dissipative and structural residuals.

  • Prior approaches: Prior task-frame methods typically assume known geometry, while newer approaches infer constraints or task-aligned frames from demonstrations using predefined models, kinematics, or power objectives.These methods aim to recover a control structure induced by task constraints.
  • Interaction-frame formulation: The interaction frame replaces geometric priors with environmental compliance, providing a physically grounded basis for recovering control structure across contact-rich tasks.The formulation uses compact approximations based on energy dissipation.
  • Interaction-frame formulation: Under locally conservative contact, environmental stiffness is decomposed into constrained high-stiffness directions and admissible low-stiffness motion directions.The formulation models the local response with a symmetric stiffness matrix and separates its principal axes into constraint and admissible-motion subspaces.
  • Frame construction: The frame uses the dominant interaction wrench to define the z-axis and projects motion intent onto the orthogonal plane to define the x-axis.Degenerate cases with zero projection use a canonical reference and complete the frame by the right-hand rule.
  • Frame recovery: The recovery procedure distinguishes dissipative residuals along intended motion from structural residuals along intended wrench and orthogonalizes the corresponding signals accordingly.The method assumes gravity and inertial effects are compensated or negligible and treats the observed twist and wrench as approximations to task intent.
  • Prior limitations: Power-minimization and structural-only approaches can miss friction-dominated interactions, lack physical interpretability, or converge to spurious frames.The paper identifies these as limitations of prior interaction-frame recovery methods.

IV. METHOD

The method recovers task control structure from demonstrations and combines it with asynchronously scheduled global and local policies. Demonstrations provide visual, state, twist, and wrench signals under gravity-compensated, smooth-motion assumptions.

  • Method overview: The method first recovers the control structure of contact-rich tasks, then introduces Force Policy and a dual-policy asynchronous scheduler for deployment.The scheduler manages dual-frequency execution to produce smooth trajectories.
  • Demonstration representation: Each demonstration records visual observations, end-effector poses and states, robot twists, and wrenches over the task horizon.The twist contains linear and angular velocity, while the wrench contains force and moment.
  • Assumptions: The method assumes gravity-compensated wrench measurements and negligible inertial effects because human demonstrations have smooth motion profiles.

A. Control Structure of Contact-Rich Tasks

The paper recovers an interaction frame from demonstrations and uses it to organize hybrid force-position control around task-specific interaction modes. Force Policy separates global vision-guided progression from high-frequency local force regulation during contact.

  • Interaction Frame Identification: The interaction frame is an instantaneous local basis that captures task-relevant interaction structure and is recovered from real-world demonstrations.The method assumes the frame origin is anchored at the end-effector and estimates its orientation from visual observations, measured signals, and task descriptions.
  • Task Classification and Control Signal Generation: Interaction patches are classified as Free, Surface, Insertion, or Rotation using interaction-frame-aligned twist and wrench characteristics.These modes distinguish negligible contact, dominant normal force, insertion friction, and rotation-dominated interaction.
  • Task Classification and Control Signal Generation: A selection mask maps each interaction mode to force- or position-controlled subspaces, enabling motion while regulating contact forces.The interaction-frame x-axis aligns with intended motion, while the z-axis aligns with the intended interaction wrench direction and dominant normal.
  • Global-Local Vision-Force Decomposition: Force Policy assigns task-level progression to a global vision policy and contact execution to a high-frequency local force policy.The global policy acts directly in free space, while the local policy uses force feedback and local observations to regulate interaction.
  • Force Policy: The local policy predicts interaction structure and force-based action parameters, including controlled subspaces, motion chunks, and reference wrenches.The global visual feature conditions local inference without requiring the global and local policies to run at the same frequency.
  • Force Policy: The predicted selection mask also serves as an implicit router that transfers control authority from global vision to hybrid force-position regulation.Non-zero force-control entries activate local force control axes as needed during interaction.

C. Dual-Policy Asynchronous Scheduler

The dual-policy asynchronous scheduler coordinates global vision and high-frequency local force inference without allowing global latency to disrupt closed-loop execution. It aligns delayed action chunks and smooths trajectories for stable, force-consistent deployment evaluation.

  • Scheduler Design: The scheduler runs global and local policies asynchronously, uses smooth routing transitions, and time-aligns delayed global outputs with execution.This design prevents global inference latency from degrading high-frequency local control.
  • Evaluation Setup: The evaluation spans three polishing and insertion tasks requiring accurate force regulation, with randomized object placement during data collection and testing.Each task is evaluated under contact-rich conditions, and demonstrations are collected with force feedback.
  • Chunk Alignment with Dynamic Time Warping: Dynamic time warping aligns predicted action trajectories with recent execution history and drops preceding waypoints to remove latency-induced discontinuities.The waypoint-dropout strategy is model-agnostic and avoids modifying model architectures or training procedures.
  • Scheduler Design: Unified-frequency resampling, acceleration-continuous planning, and jerk suppression produce smooth trajectories for hybrid force-position execution.The resulting trajectories are dispatched to a non-real-time hybrid force-position controller.
  • Evaluation Results: Force Policy achieves the highest success rates among vision-based and force-aware baselines across all evaluated tasks.Success is reported using accumulative, stage-wise evaluation, with partial completion counted as half success for two specified phases.
  • Force Behavior: On the Scrape off Sticker (Hard) task, Force Policy closely imitates demonstrated effective forces while baselines exhibit degraded force behavior and performance.The comparison links force-profile fidelity with higher success rates in this task.

B. Results

Force Policy improves contact-rich manipulation by separating stable global visual guidance from high-frequency local force control. Across diverse tasks, it provides stronger force regulation and generalization, especially during rapid contact phases and on unseen objects.

  • Force Policy consistently outperforms vision-only and force-aware baselines across diverse contact-rich tasks, with largest gains during precise, rapid force-regulation phases.The strongest improvements occur during insertion in Plug in EV Charger and scraping in Scrape off Sticker.
  • Table III evaluates force regulation on Push and Flip using pushed distance d, with approximately 45N required for heavy-object pushing and about 15N used to flip the target.The table compares pushed-distance statistics across methods under these force demands.
  • Global-local decoupling insulates long-horizon visual intent from noisy force signals while letting the local policy react independently to real-time contact feedback.This design targets robustness under sensor noise and fast contact transients.
  • Force Policy achieves superior force control by tracking human demonstration profiles and maintaining effective force magnitudes during critical phases.Baselines exhibit oscillations, insufficient downward force, or excessive force that can result in task failure.
  • On unseen objects varying in color, geometry, and stiffness, Force Policy consistently achieves high success rates while baselines frequently fail catastrophically.The evaluation tests visual generalization to unseen objects and force-regulation generalization to unseen contact geometries.

C. Ablations

The ablations examine interaction-frame recovery, semantic classification, asynchronous scheduling, and the scope of the proposed theory and evaluation. Results indicate accurate frame recovery and semantic classification, smoother scheduled trajectories, and explicit boundaries on contact settings, task coverage, and torque control.

  • Interaction-frame recovery: Angular error between recovered and ground-truth force-control axes is evaluated on Scrape off Sticker, with failures counted above 20°.The evaluation uses a 20° failure threshold.
  • Interaction-frame recovery: Twist-only and adaptive interaction-frame recovery outperform wrench-only on Scrape off Sticker, while analytic and power-based approaches show larger errors.The comparison attributes the difference to dissipative residuals such as friction dominating power in this task.
  • Semantic classification: Semantic classification accuracies reach 100.0%, 92.0%, and 98.0% across the three evaluated tasks.The authors state that infrequent misclassifications and online force-feedback correction limit their effect on policy learning.
  • Asynchronous scheduling: The asynchronous scheduler reduces motion and force jerks during Push and Flip, producing smoother executed trajectories.Trajectory smoothness is evaluated using spectral arc length, SPARC (↑).
  • Scope: The interaction-frame theory covers stable-contact manipulation under locally conservative assumptions but excludes fracture and non-conservative contact transitions.Multi-point and surface-contact cases are discussed separately in the supplementary material.
  • Scope: The evaluation covers peg-in-hole and surface-interaction tasks, while broader complex contact-rich tasks remain future work.Push and Flip tests continuously changing interaction frames.
  • Scope: The formulation focuses on recovering interaction orientation rather than explicitly estimating the true contact point, so torque control requires fuller contact modeling.The limitation concerns extending the method beyond force-control orientation recovery.

A. Geometry-Induced Interaction Structure

The paper models contact wrenches as conservative and dissipative components, then derives a geometry-induced interaction basis from environmental stiffness. Under Hertzian contact and symmetry assumptions, this basis diagonalizes stiffness and separates constrained from admissible directions.

  • Physical decomposition: The total measured wrench combines conservative elastic effects with non-conservative dissipative effects such as friction.The decomposition is used to extract geometric topology from the measured interaction response.
  • Environmental stiffness: Environmental stiffness is defined from the sensitivity of the elastic wrench to pose perturbation and is symmetric as the Hessian of a conservative potential.This symmetry supports spectral analysis of the contact structure.
  • Unified spatial basis: Under Hertzian contact, a single spatial rotation diagonalizes the full 6×6 environment-stiffness matrix.The theorem uses a unified rotation for translational and rotational stiffness.
  • Geometric frame: The principal geometric frame uses the surface normal and the contact ellipse’s principal axes, whose reflectional symmetry makes the contact structure geometrically decoupled.The basis is {t1, t2, n}, with t1 and t2 aligned to principal relative-curvature directions.
  • Diagonalization: In the principal geometric frame, the stiffness matrix is fully diagonalized by the corresponding spatial rotation.The result follows after the off-diagonal translational, rotational, and linear-angular terms vanish.
  • Generalization: The diagonalization property extends beyond Hertzian contact when the stiffness distribution and contact domain have macroscopic reflectional symmetry.The paper discusses multi-point and planar area contacts as relevant examples.
  • Spectral-geometric correspondence: The eigenbasis of environmental stiffness aligns with the principal geometric axes, mapping high-stiffness directions to constraints and near-zero stiffness directions to admissible motion.The constraint subspace U uses λ_i > ε, while the admissible-motion subspace T uses λ_i ≤ ε.

B. Task Intent and Interaction Frame

The interaction frame resolves the orientation and semantic ambiguity left by spectral stiffness analysis by anchoring principal axes to task intent. Under compatibility and symmetry assumptions, intended forces and motions induce co-axial responses in the resulting frame.

  • Task intent: Task intent is defined by the desired interaction wrench and twist at the interaction point, providing the directional anchor for the interaction frame.The frame uses intent to assign meaning to otherwise ambiguous spectral axes.
  • Task intent: The intent-compatibility assumption places forces primarily in constrained directions and motions primarily in admissible directions.This also implies orthogonality between force and motion intents because U ⊥ T.
  • Frame construction: The frame prioritizes a non-negligible wrench axis as the constraint axis; otherwise, the twist axis defines the motion axis.A secondary axis is obtained by projection, with a reference vector used when the secondary intent is degenerate or collinear.
  • Frame construction: The final axis completes an orthonormal basis by the right-hand rule, and the resulting interaction frame diagonalizes environmental stiffness.This construction resolves directional ambiguity and co-axial degeneracies.
  • Intent alignment: The interaction frame is co-axial with the principal geometric frame, so its spatial rotation diagonalizes the environmental stiffness matrix.This is the stated consequence of intent alignment under the paper’s assumptions.
  • Force response: An applied wrench aligned with a constraint axis induces an elastic pose perturbation and parasitic twist along that same axis.The theorem derives the scalar response δχ_j = β/k_j for a constraint-axis wrench.
  • Motion response: An intended twist along an admissible-motion axis produces a co-axial dissipative wrench when local dissipation and contact structure satisfy the stated reflectional symmetries.The dissipative response opposes motion under the proposition’s assumptions.
  • Approximation basis: The resulting parasitic wrench and parasitic twist remain aligned with their respective intended twist and wrench directions, establishing the basis for the paper’s approximation strategy.The stated orthogonality relation is W_c ∥ ξ* ⊥ W* ∥ ξ_c.

SUPP. II. INTERACTION STRUCTURE DISCOVERY

The paper argues that wrench and twist observations alone are ambiguous about interaction structure, so high-level knowledge is needed to recover a physically meaningful Interaction Frame. Across three tasks, the recovered frame axes align with task geometry and remain topologically stable, while wrench-only recovery is distorted by friction.

  • Interaction structure ambiguity: Wrench and twist signals can be identical for dissipative friction and structural spring interactions, so sensor observations alone cannot identify the energy destination.The paper uses pushing a heavy block against friction and compressing a stiff spring as an example of this ambiguity.
  • Interaction Frame recovery: The recovered Interaction Frame aligns its axes with task geometry across Push and Flip, Plug in EV Charger, and Scrape off Sticker demonstrations.The authors report this qualitative consistency across varying expert demonstrations.
  • Interaction Frame recovery: The recovered spectral axes are topologically stable across demonstrations of the three contact-rich tasks.This is presented as qualitative validation of the physical consistency of the recovery method.
  • Interaction Frame recovery: On Scrape off Sticker, wrench-only recovery couples friction directions, whereas the proposed method better aligns with the ground-truth world z-axis.Tangential scraping friction acts as a parasitic wrench for the baseline, distorting its inferred force direction.

SUPP. III. FORCE POLICY

Force Policy combines a low-frequency global vision policy with a high-frequency local force policy, then synchronizes their outputs for smooth execution. The local policy predicts interaction structure and high-frequency actions, while trajectory alignment and latency compensation support continuous control.

  • Policy architecture: Force Policy uses a global vision policy to provide high-level intent and geometry features that guide the local force policy.The global policy processes low-frequency global observations and exposes its latent action embedding as a 512-dimensional visual feature.
  • Policy architecture: The local force policy fuses wrist vision, proprioceptive history, and global context through adaptive gating that varies with the interaction phase.Its inputs include wrist images, end-effector pose history, and wrench history.
  • Local force policy: The interaction structure head predicts the Interaction Frame pose, reference wrench, and binary active-control-mode mask.These outputs provide task-oriented structural parameters for hybrid force-position control.
  • Local force policy: The action head predicts 50Hz relative-action chunks and executes only the first action using receding-horizon control.The local policy is trained end-to-end with action and auxiliary interaction-structure losses.
  • Execution scheduling: Dynamic Time Warping selects the trajectory index matching the current robot state, preserving position and velocity continuity while compensating for scheduler latency.The procedure takes approximately 20ms, corresponding to dropping an additional 1–2 steps at 50Hz.

B. Broader Impact on General Policies

The asynchronous scheduler improves contact-rich execution beyond Force Policy itself by coordinating inference and trajectory execution. Across vision-only policy architectures, it increases success and reduces control error, with waypoint dropout helping prevent contact loss.

  • Scheduler impact: The asynchronous scheduler overlaps inference with trajectory execution and interpolates outputs to a unified 50Hz rate.It selectively drops waypoints to mitigate action-chunk mismatch caused by inference latency.
  • Scheduler impact: For RISE-2, the scheduler increases flipping success from 42.5% to 62.5% and reduces average displacement error to zero.Pushing success remains perfect in the reported comparison.
  • Scheduler impact: The scheduler consistently boosts success rates and reduces control error across the evaluated policy architectures.The reported evaluation includes RISE-2 and the VLA model π0.5.
  • Mechanism: DTW-based waypoint dropout helps prevent contact loss caused by temporally inconsistent action chunks and erroneous retreating motions.Such retreating actions can momentarily detach the end-effector from the object surface during contact-rich flipping.

E. Failure Analyses

Failure analyses show that baselines commonly lose contact, transition at the wrong time, or apply insufficient force, especially under tight force and geometry requirements. Force Policy remains robust across disturbances and unseen object variations, although connector insertion can still approach hardware limits.

  • Failure modes: Most Push and Flip failures occur when policies attempt to flip before establishing wall contact, making contact-then-transition logic difficult to learn from demonstrations.The transition occurs within roughly 50ms in human executions, and vision-only policies are especially affected.
  • Failure modes: Position-control force-aware methods often fail to maintain stable contact during flipping, causing the box to slip or drop.Force Policy instead uses force control to establish and maintain contact.
  • Task-specific limitations: Plug insertion remains challenging because the required force is near the arm’s torque limits, and small motion or Interaction Frame errors can trigger hardware protection.The paper states that Force Policy can also encounter this failure mode and may require finer-grained force regulation.
  • Generalization: Force Policy generalizes to unseen objects varying in geometry, stiffness, material compliance, color, and texture.The authors attribute this result to explicit Interaction Frame modeling and adaptive proprioceptive fusion.
  • Contact robustness: During human disturbances, Force Policy adapts end-effector pose to maintain the target wrench, keeping contact force smooth and constant.The policy prioritizes force stability over geometric strictness during contact maintenance.
  • Final-stage results: Final-stage results with 95% confidence intervals show that Force Policy significantly outperforms the baselines.The reported table presents final-stage success rates in percent.
Loading 2602.22088v2…