Source-linked AI summary
A Survey of Robot Manipulation in Contact
Markku Suomalainen, Yiannis Karayiannidis, Ville Kyrki
TL;DR
Robotic manipulation in contact covers force-aware tasks that can exploit environmental interaction to address uncertainty, including tasks traditionally performed by humans. This survey synthesizes the task landscape, control and policy representations, and planning and learning methods, while identifying continuing limitations in learning speed, generalization, simulation transfer, and exception handling.
Problem
Robots need to perform contact-rich tasks requiring explicit or implicit interaction-force control, including tasks where contact helps mitigate uncertainty.
Method
The survey organizes state-of-the-art methods by in-contact tasks, controllers, policy representations, and planning or learning approaches.
Results
The surveyed literature shows robots performing increasingly human-associated contact tasks, while classical tasks such as peg-in-hole receive improved generalization, error tolerance, and planning or learning methods.
Takeaways & Limitations
Contact is increasingly treated as a useful resource for factory and service robots rather than only as a nuisance.
Takeaways & Limitations
The field still needs faster learning and better generalization, improved simulation-to-reality transfer, and stronger handling of atypical errors and complex tasks such as painting.
Abstract
from arXiv · showhide
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 frames manipulation in contact as force-aware robot tasks spanning continuous environmental interaction, from assembly to human-scale service skills. It organizes the field around tasks, controllers, policy representations, and planning or learning methods.
- Motivation: Contact is increasingly exploited to localize robots and tools under uncertainty rather than treated only as a nuisance.The survey notes that methods for managing and exploiting contact have evolved while keeping computational demands feasible.
- Scope: Manipulation in contact requires explicit or implicit control of interaction forces during extended contact between the robot, its hand, tool, and environment.The survey includes tasks ranging from screwing and peg-in-hole to excavation, wood planing, melon scooping, and massaging.
- Controllers: Explicit force control sets desired force levels, commonly through force or hybrid force/position controllers with high-frequency updates.A classical force controller uses a PI loop to maintain the applied force near a desired level.
- Controllers: Implicit force control is commonly implemented as compliance, including software impedance control and, in selected works, mechanical compliance.Impedance control permits deviation from a desired trajectory, supporting both free-space and in-contact motion without switching controllers.
- Representations and learning: The survey separates higher-level skill conveyance into planning, Learning from Demonstration, and Reinforcement Learning after discussing policy representations.It also distinguishes this focus from broader surveys on robot learning, reinforcement learning, learning from demonstration, assembly, and vision-based manipulation.
2 Tasks requiring manipulation in contact
The survey covers tasks in which force control is required, beneficial, or useful for handling uncertainty, including shaping materials, aligning workpieces, articulated motion, and tool-mediated interaction. Across these tasks, robots use explicit or implicit force control, compliance, contact-state estimation, and planning to manage varied physical constraints.
- Task definition: Manipulation in contact includes tasks requiring explicit or implicit interaction-force control, while some tasks benefit from force control despite being possible without it under perfect knowledge.Peg-in-hole and similar alignment tasks exemplify contact used to mitigate uncertainty.
- Tool-mediated tasks: Tool use introduces tooltip-location uncertainty when the tool is grasped, increasing the need for compliance during contact.The survey distinguishes neither rigidly attached nor grasped tools, except for this uncertainty effect.
- Environment shaping: Wiping and polishing mainly require maintaining surface contact, whereas scraping, wood planing, deburring, sanding, and engraving remove material and generally employ explicit force control.The survey characterizes wiping and polishing as compatible with implicit or explicit force control, while material-removal tasks are described as explicit-force tasks.
- Environment shaping: Excavation is challenging because hydraulic manipulators have nontrivial closed-loop interactions and excavated material varies in resistance.Autonomous excavation has been accomplished without force feedback, while other work estimates forces or applies impedance control.
- Workpiece alignment: Workpiece alignment spans peg-in-hole variants in factories and homes, including low-clearance, multi-peg, dual-arm, threaded, articulated, and elastic-piece tasks.Compliance has a long history in these tasks, and both explicit and implicit force control remain in use.
- Workpiece alignment: Humans still outperform robots in peg-in-hole generalization, surprising situations, uncertainties, and very tight clearances, although research targets these gaps.Reported approaches include meta-reinforcement learning for generalization and clearances smaller than robot accuracy, including 6µm.
- Articulated motions: Articulated-object manipulation requires interaction to perceive constrained kinematic structure, and door-opening methods estimate motion direction while using compliant behavior.The surveyed approaches commonly model everyday articulated objects as one-degree-of-freedom mechanisms.
3 Control of manipulation in contact
The survey organizes contact control around explicit versus implicit force regulation, including direct, hybrid, parallel, stiffness, impedance, and admittance approaches. These controllers differ in how they divide force and motion objectives, handle free-space and contact motion, and depend on environmental knowledge or measurements.
- Force-control categories: The survey distinguishes direct force control from controllers that regulate dynamic relationships between forces and displacements, including impedance and admittance control.
- Force-control categories: Hybrid controllers explicitly decompose force and motion into constrained subspaces, whereas parallel controllers use separate loops without constraint-defined subspaces.
- Force-control categories: Hybrid and parallel force/position structures can minimize position error without knowing contact compliance or plane position, but require the surface normal vector.
- Stiffness control: Stiffness control applies to both contact and free-space trajectories, while hybrid and parallel force controllers are unsuitable for free-space motion because they target nonzero contact force.
- Impedance control: Impedance control follows a trajectory while allowing displacement proportional to gains, modeling the interaction as a virtual spring-damper-mass system.
- Impedance control: Impedance controllers may become unstable because virtual springs store energy, especially with learning or non-stiff environments; proposed remedies include energy dissipation, energy tanks, and filtering.
4 Representing manipulation in contact
The survey organizes manipulation-in-contact skills around representations that map task requirements and sensor feedback to controller actions. It distinguishes continuous policies, which output motion or force commands, from discrete policies, which select among subpolicies, and reviews probabilistic, dynamical-system, and neural representations.
- A policy maps task requirements and sensor feedback to actions that command a low-level controller to apply forces or move the robot.
- Hierarchical policies can select among subpolicies, with continuous representations producing motion or force commands and discrete representations choosing subpolicies.
- Continuous representations: DMP-based motion primitives are used mainly with learning from demonstration and also with reinforcement learning after human demonstrations.
- Continuous representations: DMP extensions support in-contact tasks through explicit force control, admittance or impedance profiles, and task parameterization for variations such as peg-in-hole depth.
- Continuous representations: GMM-based methods encode trajectories with multidimensional Gaussian distributions and differ mainly in how transitions between distributions are handled.
- Continuous representations: SEDS, UMIC, neural networks, and other dynamical or function-approximation methods extend representation choices for in-contact skills, with neural networks encoding policies or trajectories.The survey notes that learning environment dynamics with deep networks is seldom used because collecting sufficient training data is difficult.
- Discrete representations: HMM-based discrete representations infer hidden states from observations and can segment demonstrations or select suitable policies during execution.Autoregressive extensions support tasks involving constant or varying contact, including assembly, slicing, and grating.
5 Learning and planning of manipulation in contact
The survey reviews planning, reinforcement learning, and learning from demonstration as ways to find policies for contact-rich manipulation. These approaches incorporate contact states, force or impedance profiles, demonstrations, simulation, and safety mechanisms to address uncertainty and contact hazards.
- Finding a policy typically means determining parameters of a chosen representation, and planning and learning are often combined or use the same representations.
- 5.1 Planning: In-contact planning must represent not only robot motion but also force or impedance profiles because the robot deliberately interacts with obstacles.The survey describes force-contact sequences for localization and impedance profiles for compliant motions that overcome contact uncertainty.
- 5.1 Planning: Planning methods include force-based peg-in-hole assembly, compliant-motion planning, probabilistic roadmaps, MDPs, POMDPs, and feedback MDPs.These methods address actuation uncertainty, contact-state switching, and efficient planning when CAD or mesh models are available.
- 5.1 Planning: Contact can localize the robot or workpiece and mitigate uncertainty through contact-state detection, filtering, Bayesian estimation, fixture placement, and physical exploration.
- 5.2 Reinforcement learning: Reinforcement learning has been applied to in-contact tasks while optimizing set-point trajectories, controller parameters, or complete feedback policies.
- 5.2 Reinforcement learning: Physical reinforcement-learning exploration is hazardous under potentially high contact forces, motivating torque control, limited-stiffness impedance control, and action limits.
- 5.2 Reinforcement learning: Simulation enables safer exploration and abundant training data, but reality gaps require calibration, noise, domain randomization, or real-world refinement.Physical experiments remain necessary to evaluate closed-loop performance because simulated reality gaps differ from physical ones.
- 5.3 Learning from demonstration: Learning from demonstration uses human task executions and can be paired with reinforcement learning, optimization, or classical planning for further policy improvement.
6 Conclusions and future work
The survey synthesizes state-of-the-art methods for robot manipulation in contact and identifies important gaps in flexible-object handling, exception strategies, painting, and learning efficiency.
- The survey organizes in-contact robotics around tasks, control, representations, and methods for exploiting skills correctly.It positions these capabilities as a step toward increased automation in factories and homes.
- In-contact manipulation is relevant to factory assembly and service settings, where compliance can support varied products and expand robot abilities.
- Most surveyed tasks use stiff objects, while relatively few works address bending objects and the complex interaction between deformation and friction.
- Exception handling remains underdeveloped, leaving robots less capable than people when non-typical errors occur in industrial use cases.
- Future work should improve learning time, generalization, physics simulation, and transfer for difficult contact-and-friction tasks.The survey specifically highlights one-shot imitation learning, transfer learning, and better simulation as possible directions.