Peg-in-hole insertion

Also called: peg insertion, peg-hole assembly
Category: Planning & manipulation › Manipulation
Tags: assembly, contact-rich-manipulation, force-control, compliance, benchmark

Inserting a part into a closely fitting hole; the standard benchmark for contact-rich assembly, where the clearance is often smaller than the robot's positioning error.

Definition

Peg-in-hole insertion is the assembly task of pushing a part, the peg, into a mating feature, the hole, with a small clearance between them. It stands in for a large share of real assembly operations: shafts into bearings, pins into bores, connectors into sockets and fasteners into holes. Because the task is easy to state but hard to do reliably, it has served for decades as the standard test for robot assembly and force control.

The difficulty is that clearances are usually smaller than the combined error of the robot's positioning, calibration and perception. A purely position-controlled robot that is slightly off either misses the hole or presses hard against its edge. Reliable insertion therefore uses contact: the robot moves until it touches, uses the measured forces and moments to infer how it is misaligned, and corrects. Classical quasi-static analysis, notably by Daniel Whitney, describes the two main failure modes. Jamming occurs when the forces and moments applied to the peg are in the wrong proportion to move it; wedging occurs when the peg touches the hole at two points and friction locks it in place. Chamfers on the peg or hole and compliance at the wrist help guide the peg in. The remote center compliance device, developed at the Charles Stark Draper Laboratory, is a passive wrist that places its center of compliance near the peg tip, so that contact forces correct lateral and angular errors on their own.

A typical active strategy splits the task into phases: approach; search for the hole with a spiral or tilted motion while pressing lightly on the surface; align once the peg drops partly into the hole; and insert under force or impedance control while watching for jams. Vision usually handles coarse alignment and force sensing handles the final fraction of a millimeter. Learning-based methods, including reinforcement learning and residual policies added to a hand-designed controller, are often benchmarked on peg-in-hole, as are category-level methods that must insert parts of varying shape without a model of each one. Connectors and snap-fit parts with near-zero clearance add the need to push through interference while staying aligned.

Why it matters

Insertion is one of the most common operations in assembly and one of the hardest to automate without precise fixtures and calibration. Methods that handle it reliably under uncertainty are a key step toward flexible assembly cells that can switch parts without being reengineered.

Key applications

Fitting pins, shafts and bearings in mechanical assembly
Plugging cables and connectors in electronics assembly and charging
Benchmarking force control and contact-rich learning methods
Fastener and screw insertion
Category-level manipulation with varying part geometry

In practice

An integrator automating the fitting of steel pins into machined bores has a robot whose positioning error after calibration is larger than the bore clearance. Rather than chase more accuracy, they mount a six-axis force-torque sensor at the wrist and run an impedance controller with low lateral stiffness. A camera locates the part to within a few millimeters. The robot approaches above the hole, presses down lightly and traces a small spiral; when the downward force drops and the measured moment shows the pin has started to enter, it switches to an insertion phase that pushes down while letting the wrist comply sideways. If the axial force rises above a threshold before the target depth is reached, the controller treats it as a jam, backs off slightly, and retries with a small tilt correction based on the measured moments.

Common misconceptions

Better calibration makes force sensing unnecessary — tight clearances are usually below what a standard robot and camera can achieve, so contact-based correction is still needed
Pushing harder frees a stuck peg — jamming and wedging come from the direction of forces and friction at the contacts, and extra force usually makes them worse
Peg-in-hole is a solved toy problem — the simple geometry hides real difficulty with tight clearances, flexible parts, varied materials and unknown part shapes
Compliance has to come from the controller — passive wrist compliance, chamfers on the parts and fixture design can do much of the work without active force control

Who should know this

Assembly and automation engineers
Force-control and manipulation researchers
Robot learning researchers working on contact-rich tasks
Designers of parts and fixtures for automated assembly

Related technologies

Remote Center Compliance Device
Six-Axis Force-Torque Sensor
Spiral Search Strategies
Contact-Rich Reinforcement Learning
Chamfered Part Design

Related terms

Related projects

MuJoCo: DeepMind physics engine for robotics and ML research
ManiSkill: GPU-parallelized simulation for generalizable manipulation skills
robosuite: Modular MuJoCo simulation framework for robot learning research
LIBERO: Lifelong robot learning benchmark with 130 procedural tasks
Gymnasium-Robotics: Robotics RL benchmark environments for Gymnasium, from the Farama Foundation
MetaWorld: Multi-task and meta RL benchmark for robotic manipulation