A Survey of Robot Manipulation in Contact

In plain words

Many everyday robot jobs are not just moving to a point. Screwing something in, wiping a table, sanding a board, or opening a door means the robot has to push on something and keep the push at a sensible level. Too little force and the job fails; too much and the part or the robot is damaged. This survey collects the work done on those jobs.

It sorts the work into layers. At the bottom are controllers that decide how hard the robot pushes or how stiff it feels, such as a spring-like behaviour for the gripper. Above them are ways to describe a skill, such as a recorded motion that is bent by a force profile, a probabilistic model of several demonstrations, or a set of states that switch when contact is detected. At the top are the ways a skill is found: planned by hand from a model of the task, learned by trial and error, or copied from a person who demonstrates it.

The authors note that the field is moving from treating contact as a problem to using it, for example to find a hole by touch. They also list open problems, such as objects that bend, fast learning from few examples, and better simulators.

TL;DR

A survey, not a method. It covers how robots perform tasks that need controlled contact, from the task list through controllers and policy representations to planning, reinforcement learning and learning from demonstration §1. It reports no experiments and no results of its own, so nothing here is a benchmark number. Use it as a map of the literature and check each cited method in its own paper before implementing it.

Problem

The survey defines manipulation in contact as tasks that need explicit or implicit control of the interaction forces, where the robot or its tool stays in contact for long periods §1. Dynamic tasks such as throwing or batting, purely vision-based manipulation, and mechanical compliance (variable-stiffness actuators, soft robots) are excluded §1. Only works that explicitly show performance on an in-contact task are included §1.

The stated reason this matters: uncertainty in tasks such as peg-in-hole means the robot needs contact control to avoid excessive collisions, and compliance can let a robot insert a peg with less clearance than its own positioning accuracy §2.

Method

The survey has no method of its own. Its structure, which is the only thing an implementer can take from it, is summarised after the pseudocode below.

The one equation an implementer may need is the impedance law. Solving the virtual spring-damper-mass model eq. 1 for the commanded acceleration gives the form below. This rearrangement is ours; the survey only says an acceleration command is generated from the force and state feedback §3.2.

# Impedance law rearranged from [eq. 1]: commanded acceleration
# Inputs: position p, velocity v, desired p_d, v_d, a_d, measured force f
# Desired apparent inertia M, damping K_D, stiffness K_P (not stated)

def impedance_acceleration(p, v, p_d, v_d, a_d, f, M, K_D, K_P):
    # Tracking error terms weighted by damping and stiffness  [eq. 1]
    tracking = K_D @ (v - v_d) + K_P @ (p - p_d)    # (3,)
    # Solve M (a - a_d) + tracking = -f for a  [eq. 1]
    accel = a_d - inverse(M) @ (tracking + f)       # (3,)
    return accel

1. Tasks §2. Tasks are grouped into environment shaping, workpiece alignment, articulated motions and dual-arm manipulation §2.1 §2.2 §2.3 §2.4. Environment shaping covers wiping and polishing, grinding, planing, scraping, sanding, engraving, scooping and excavation §2.1. Workpiece alignment is mainly peg-in-hole and its variants §2.2. Articulated motions are door, drawer and valve opening §2.3. A summary of the tasks in the included papers is in tab. 1 and examples are shown in fig. 2.

2. Control §3. The survey divides force control into direct and indirect control, and into explicit (torque-resolved) and implicit (velocity-resolved) forms §3. Its main table gives hybrid, parallel and stiffness controllers in both forms tab. 2. Hybrid and parallel controllers need the surface normal, while a stiffness controller also works in free space §3.1. Impedance control makes the robot behave like a virtual spring-damper-mass system around a desired trajectory §3.2 eq. 1, with the block structure in fig. 3. Admittance control instead lets a measured force shift a reference that feeds an inner motion controller §3.3 fig. 3.

3. Representations §4. A policy maps sensor input to a control action, and can be hierarchical with a discrete choice between sub-policies on top §4. Continuous representations are:

• dynamic movement primitives (DMP), whose transformation system is a spring-damper driven by a forcing function, with a canonical phase variable §4.1 eq. 2; in contact work they are combined with explicit force or admittance control, or augmented with a force or impedance profile §4.1;
• Gaussian mixture models with Gaussian mixture regression, and the dynamical-system variant SEDS §4.1 fig. 4;
• neural networks, either as the policy or as the trajectory for a fixed low-level controller §4.1;
• UMIC, a combined motion and variable-impedance dynamical system, and the LMC representation for workpiece alignment §4.1.

Discrete representations are threshold rules on contact forces, hidden Markov models and their extensions (including autoregressive and hidden semi-Markov forms) §4.2 fig. 5 fig. 6, and Markov decision processes and partially observable ones §4.2 fig. 7. Exception strategies after a failed insertion are also covered §4.2. The most common representations are listed in tab. 3.

4. Finding the policy §5. Methods are grouped as planning, reinforcement learning, or learning from demonstration, and tab. 4 assigns the surveyed papers to these groups §5. Planning covers force-threshold planning, contact localisation, and compliant motion planning, including preimage and feedback MDP approaches §5.1. Reinforcement learning covers real-robot learning with safety limits, simulation with system identification, noise or domain randomisation, and RL that tunes controller gains or set-points §5.2. Learning from demonstration covers how to record force data (force-torque sensor, joint torques or currents, hydraulic pressures, EMG), how to align multiple demonstrations (for example with dynamic time warping), and how to pair demonstrations with later RL or planning §5.3.

Hyperparameters

The survey gives no experimental hyperparameters. The controller symbols in tab. 2 and eq. 1 have no numeric values in the text.

Results

Not applicable. The survey reports no experiments and no measured numbers. Its quantitative content is only the assignment of cited papers to categories in tab. 1, tab. 2, tab. 3 and tab. 4.

Paper vs code

No official code. The arXiv metadata and the LaTeX source contain no repository link or project page. A GitHub repository search for the title words returned nothing, and a search on the author surname returned unrelated repositories. A survey has no implementation of its own, so none is expected.

Pitfalls

• Nothing to implement. The card's Method section describes a literature map, not an algorithm. Do not expect an equation or procedure from the survey that is not in a cited paper.
• Table rows are cited claims, not verified results. For example, a paper is listed under "RL" in tab. 4 or under a task in tab. 1 because the survey placed it there. Check the cited paper for what it did.
• Some categories are judgement calls. The authors say the policy-finding table only covers works where finding the policy was a major contribution §5. Other papers are only mentioned in the text.
• Sign and notation in the controller table. eq. 1 writes the right-hand side as minus the force, but the text does not define the sign convention of the measured force. tab. 2 is for a non-redundant robot with a 3x3 Jacobian and a surface normal, and the caption says the controllers ignore orientation and torque tab. 2 §3.1.
• The DMP equation is a summary. eq. 2 is written with the text's own symbols, and the canonical function is given as a decaying phase with time constant tied to the demonstration length §4.1. Consult the original DMP papers cited in the survey for the exact formulation.
• Scope gaps the authors name. Deformable or bending objects, painting, and exception handling in industrial use are described as open §6. Dynamic tasks and vision-based manipulation are out of scope §1.
• Commented-out material. The LaTeX source contains commented-out tables, equations and figures. They are not part of the survey text and this card does not use them.
• Version. The metadata gives arXiv version v3 (2022-07-25). The venue is not stated in the source, and later versions may differ.

Lineage

• Related surveys named by the authors: a compliance-control survey from 2014 §1; a survey of non-prehensile (dynamic) manipulation from 2018 §1; and surveys of learning from demonstration and of reinforcement learning for robotics §1. This card does not link them because they have no cards here.
• Compared against: not stated. The survey is not a benchmark.
• Superseded by: not stated.