PaperlayerView exact revision on arXiv

Source-linked AI summary

A Survey of Robot Manipulation in Contact

Markku Suomalainen, Yiannis Karayiannidis, Ville Kyrki

arXiv:2112.01942v3cs.RO

TL;DR

Robots increasingly perform contact-rich manipulation, but the field spans diverse tasks, control strategies, representations, and learning or planning methods. This survey organizes the current state across these dimensions and highlights progress toward robots exploiting contact in assembly and service tasks.

  • Problem

    Contact-rich manipulation requires robots to control interaction forces across diverse tasks, while the field lacks a unified overview of these tasks and approaches.

    Source: §1 · p. 2

  • Method

    The survey synthesizes in-contact tasks, control methods, task representations, and learning and planning techniques for the required skills.

    Source: §3 · p. 11, §6 · p. 29

  • Results

    The survey reports state-of-the-art methods spanning traditional assembly, material removal, and human-oriented tasks, including near-optimal compliant-motion planning.

    Source: §1 · p. 2, §5.1 · p. 25

  • Takeaways & Limitations

    Contact is increasingly treated as a useful resource for expanding robotic capabilities in factories, homes, restaurants, and other service environments.

    Source: §1 · p. 2, §6 · p. 29

    Learning time and generalization remain insufficient, and improved simulation and transfer methods are still needed for complex in-contact tasks.

    Source: §6 · p. 30

Abstract

from arXiv · show

In this survey, we present the current status on robots performing manipulation tasks that require varying contact with the environment, such that the robot must either implicitly or explicitly control the contact force with the environment to complete the task. Robots can perform more and more manipulation tasks that are still done by humans, and there is a growing number of publications on the topics of 1) performing tasks that always require contact and 2) mitigating uncertainty by leveraging the environment in tasks that, under perfect information, could be performed without contact. The recent trends have seen robots perform tasks earlier left for humans, such as massage, and in the classical tasks, such as peg-in-hole, there is a more efficient generalization to other similar tasks, better error tolerance, and faster planning or learning of the tasks. Thus, in this survey we cover the current stage of robots performing such tasks, starting from surveying all the different in-contact tasks robots can perform, observing how these tasks are controlled and represented, and finally presenting the learning and planning of the skills required to complete these tasks.

1 Introduction

The survey examines robots performing extended-contact manipulation, where contact forces are explicitly or implicitly controlled and increasingly exploited to handle uncertainty. It organizes the field around task types, controllers, policy representations, and planning or learning methods, while excluding short-contact dynamic and vision-based manipulation.

  • Robots increasingly perform contact-rich tasks once considered too difficult, while exploiting environmental contact for localization and tool use under uncertainty.The survey notes that improved contact-management methods have made the computational burden feasible.Source: §1 · p. 2
  • The survey defines manipulation in contact as extended robot–environment or tool–environment contact requiring explicit or implicit force control, spanning assembly, material removal, and human-centric tasks.Dynamic tasks with short contact times, such as throwing or batting, are excluded.Source: §1 · p. 2
  • Contact-rich manipulation requires low-level force regulation through explicit force or hybrid force/position control, or implicit compliance via impedance control and, in some cases, mechanical compliance.The survey develops advanced impedance controllers in Section 3 while treating mechanical compliance selectively.Source: §1 · pp. 2–3
  • The survey covers higher-level and potentially hierarchical policy representations, then organizes skill acquisition into planning, Learning from Demonstration, and Reinforcement Learning.Exact planning is difficult for in-contact tasks, while LfD transfers human skills and RL searches for solutions.Source: §1 · p. 3
  • The survey focuses on works demonstrating in-contact performance with force as the sole feedback, excluding vision-based manipulation and broadly applicable methods without such demonstrations [13].Its structure proceeds from tasks to controllers, policy representations, and planning or learning methods.Source: §1 · p. 3, Figure 1 · p. 4

2 Tasks requiring manipulation in contact

The survey defines in-contact manipulation as tasks requiring explicit or implicit control of interaction forces, spanning environment shaping, workpiece alignment, and articulated motions. It also considers tool-induced uncertainty, compliance, and dual-arm variants across these task categories.

2 Tasks requiring manipulation in contact

  • In-contact manipulation is required when force interactions must be controlled explicitly or implicitly, including force-critical tasks, contact-enhanced tasks, and uncertain alignment or articulated motions.Compliance can prevent excessive collisions and enable peg-in-hole assembly with clearance smaller than robot accuracy.Source: §2 · p. 4

2.1 Environment shaping

  • Environment-shaping tasks range from wiping and polishing to scraping, planing, deburring, sanding, engraving, grinding, drilling, scooping, excavation, and velcro peeling, with force-control demands varying by precision, periodicity, and material.Implicit control may suffice for surface following, whereas material removal and massage generally require explicit or more accurate force control; excavation remains challenging because hydraulic dynamics and material resistance vary.Source: §2.1 · p. 7

2.2 Workpiece alignment

  • Workpiece alignment centers on peg-in-hole assembly and its industrial, household, multi-peg, dual-arm, threaded, elastic, folding, and snapping variants, where compliance and force control improve robustness to uncertainty and tight clearances.Robots still lag humans in generalization and surprising situations, but methods have achieved clearances smaller than robot accuracy (6µm), near-optimal compliant plans, and improved simulation for transfer learning.Source: §2.2 · pp. 8–9

2.3 Articulated motions

  • Articulated motions require interaction to infer constrained kinematics, commonly using compliant or hybrid force/motion control for doors, drawers, handles, jars, and valves.Simple mechanisms can use proprioception and force measurements, while multi-body objects require richer visual and probabilistic perception to handle noise, uncertainty, and contact loss.Source: §2.3 · pp. 9–10

2.4 Dual-arm manipulation in contact

  • Dual-arm contact manipulation often uses one arm as a fixture, but coordinated motion or compliance from both arms can improve articulated rotation, scooping, drawing, peg-in-hole, and bimanual folding tasks.For 2-DOF articulated objects, force/torque sensing and proprioception from both arms can localize joints and track their states.Source: §2.4 · p. 10

3 Control of manipulation in contact

Contact manipulation control is organized around direct or indirect force objectives and explicit or implicit control signals, with higher-level policies relying on force, impedance, and admittance methods. These methods address force regulation, compliant motion, and force-to-motion relationships while accounting for sensing, environment compliance, and stability.

3 Control of manipulation in contact

  • The survey classifies contact-force control as direct or indirect and as explicit or implicit, focusing on its use with higher-level manipulation policies.Direct control targets forces, whereas indirect control targets dynamic force–displacement relationships; control is the lowest level of a hierarchical policy.Source: §3 · p. 11

3.1 Force Control

  • Direct force control uses feedback to maintain a desired force, while hybrid and parallel force/motion controllers combine force and position objectives under different constraint assumptions.Hybrid control decomposes control subspaces using motion constraints, whereas parallel control uses separate force and compliance-based position loops; both require the surface normal and can accommodate deforming environments without knowing contact compliance.Source: §3.1 · pp. 11–12, Table 2 · p. 13

3.2 Impedance control

  • Impedance control regulates a virtual spring–damper–mass relationship rather than force directly, but varying impedance can cause instability, particularly with learning and non-stiff environments.Solutions include dissipating or storing energy, filtering impedance profiles, or using feed-forward control; complex interactions such as food cutting may require learning-based MPC instead of a linear impedance model.Source: §3.2 · p. 14

3.3 Admittance control

  • Admittance control converts measured external force into a desired position or velocity for an inner motion loop, enabling compliant manipulation without necessarily requiring torque-controlled robots.It has been applied to compliant motion, force-guided assembly, excavation, peg-in-hole, and physical human–robot interaction, including variable and nonlinear admittance structures.Source: §3.2 · p. 15, §3.3 · p. 16

4 Representing manipulation in contact

Manipulation skills are represented as policies mapping task requirements and sensor feedback to controller actions, with hierarchical forms selecting among continuous subpolicies. The survey distinguishes continuous representations for motion and force generation from discrete representations for switching skills under changing contact conditions.

4 Representing manipulation in contact

  • Policies map sensor input and state to actions for a lower-level controller, while hierarchical policies select subpolicies before issuing motion or force commands.Although some methods omit force sensing, surveyed applications generally use proprioceptive feedback such as joint encoders, so they are not completely feed-forward.Source: §4 · pp. 16–17, §4.1 · p. 17

4.1 Continuous representations

  • Continuous representations include DMPs, GMM-GMR, dynamical systems, neural networks, and specialized compliant primitives that encode motion, force, impedance, uncertainty, or task parameters for in-contact skills.DMPs and GMM-GMR require force or compliance augmentation for contact tasks, while task-parameterized CMPs can generalize variations such as peg-in-hole depth; other methods include SEDS, UMIC, LMC, and SSCP.Source: §4.1 · pp. 18–20

4.2 Discrete representations

  • Discrete representations choose among continuous skills or policies using contact phases, force thresholds, probabilistic state models, or decision processes.HMM-based methods segment demonstrations and select policies under uncertain observations, while POMDPs support planning, learning, and contact-based localization in limited in-contact applications.Source: §4.2 · pp. 20–22
  • Discrete exception strategies detect failures or stuck workpieces and switch to recovery, search, assistance, or alternative control policies during assembly.Reported strategies include HMMs, SVMs, random walks, Archimedean spirals, incremental human assistance, area-coverage methods, ergodic control, PCA, and neural-network error detection.Source: §4.2 · p. 23

5 Learning and planning of manipulation in contact

Learning and planning in contact manipulation find policies over motion, force, or impedance representations, with planning, reinforcement learning, and learning from demonstration addressing complementary aspects of skill acquisition. Contact-specific methods exploit force interaction for compliant motion, uncertainty mitigation, and policy optimization, while physical safety, simulation-to-real transfer, and force-aware demonstrations remain central challenges.

5 Learning and planning of manipulation in contact

  • Contact-manipulation policies combine motions with force or impedance profiles, using either contact sequences for localization or compliant motions to overcome uncertainty.Planning and learning may overlap and can use multiple task representations; Table 4 categorizes surveyed methods by how policies are found.Source: §5 · pp. 23–24, Table 4 · p. 24

5.1 Planning of manipulation in contact

  • Planning methods exploit contact states, force thresholds, fixtures, and compliance to solve assembly and other in-contact tasks, including peg-in-hole, folding, snapping, and excavation.Near-optimal planning has been obtained with MDPs, POMDPs, probabilistic roadmaps, feedback MDPs, and known CAD or mesh models, while contact localization uses force-position mappings, filtering, and Bayesian estimation.Source: §5.1 · pp. 25–26

5.2 Reinforcement learning of manipulation in contact

  • Reinforcement learning covers pushing, articulated-object manipulation, tool use, and assembly, but physical exploration must limit contact forces and simulation-trained policies must address the reality gap.Safety measures include torque or impedance control and action constraints; simulation transfer uses system identification, noise, domain randomization, and real-world refinement, while real experiments remain necessary for closed-loop evaluation.Source: §5.2 · pp. 26–27

5.3 Learning from demonstration of manipulation in contact

  • Learning from demonstration requires measuring or estimating contact force and selecting representations, demonstration counts, alignment, segmentation, and environment-exploiting behaviors appropriate to the task.Kinesthetic teaching often requires an F/T sensor, while teleoperation can use currents, torques, joint sensors, or hydraulic pressures; explicit instruction, incremental learning, and corrections help address inadequate demonstrations.Source: §5.3 · pp. 27–29
  • Pairing demonstrations with reinforcement learning or optimization can further improve policies when tasks require more optimization than rapid initial learning.Demonstrations can also enhance classical motion-planning methods, although whether human demonstrations are optimal remains unresolved.Source: §5.3 · p. 29

6 Conclusions and future work

The survey synthesizes state-of-the-art methods for robot in-contact tasks, from task types and control to representations and their exploitation. It identifies industrial and service applications while outlining future needs in deformable-object manipulation, exception handling, learning, simulation, and force inference.

  • The survey covers in-contact task types, robot control, task representations, and methods for exploiting those representations to perform tasks correctly.These capabilities are positioned as a path toward greater automation in factories and homes.Source: §6 · p. 29
  • In-contact manipulation can expand factory automation for varying products and give future service robots more abilities in homes, restaurants, and similar environments.The survey argues that contact should be treated as useful rather than merely a nuisance, while substantial progress remains necessary.Source: §6 · p. 29
  • Future work should address deformable-object tasks, especially assemblies where deformation and friction create complex physics beyond the predominantly stiff-object tasks studied so far.Only a few works address bending objects, including, while cloth folding has been studied in.Source: §6 · pp. 29–30
  • Robots need better exception handling, force-sensitive painting, and contact-based accuracy for stiff-object assemblies such as board-to-board matings.Manual intervention for atypical errors remains especially problematic in industrial use cases, while painting requires repeatable delicate force control.Source: §6 · p. 30
  • Future learning research should reduce teaching time and improve generalization through one-shot and transfer learning, stronger simulation and VR demonstrations, HRI guidance, and vision-based force inference beyond grasping.Relevant directions include dynamics domain randomization, accurate simulators, VR demonstrations, and learning from video for tasks such as sawing.Source: §6 · p. 30

Related papers

Loading 2112.01942v3…