Mobile manipulation

A fixed robot arm works within the reach of its base. Put the arm on a mobile base and its workspace becomes a whole building: the robot can fetch items from shelves, tend several machines, restock store aisles, open doors and help people at home. Mobile manipulation combines everything in this program, from localization and navigation (ROB-340) and motion planning (ROB-350) to grasping and force control, and it adds its own difficulties: the base is never positioned perfectly, the arm and base must cooperate, and the robot works among people in spaces not designed for robots.

This lesson covers the architectures of mobile manipulators, how base and arm motion are coordinated, the error budget that decides whether a grasp will succeed after driving to it, perception strategies that correct base errors, and the safety requirements that apply when a moving base carries a moving arm.

Kinds of mobile manipulators

Mobile manipulators come in several forms. Industrial autonomous mobile robots with a collaborative arm on top tend machines, move parts between stations and handle totes, often docking at each station before the arm works. Warehouse picking robots combine a mobile base with a lift and a gripper or suction head to pick items from shelves. Service and research robots have one or two arms on a wheeled or legged base, sometimes with a torso that raises and lowers. Legged robots with an arm on the body can climb stairs and step over obstacles. Humanoid robots, an area of intense development at the time of writing, aim to use human environments and tools directly.

The base type shapes everything. A differential-drive base cannot move sideways, so positioning in front of a shelf often requires a maneuver. An omnidirectional base with mecanum or swerve wheels can move in any direction, which simplifies fine positioning. Legged bases handle rough terrain but move the arm's mounting point with every step.

Sequential versus whole-body motion

The simplest strategy is sequential: drive the base to a parking pose near the target, stop, then move the arm while the base stays still. It is easy to plan, safe and predictable, and most industrial mobile manipulators work this way. Its costs are time, because the arm waits while the base drives and vice versa, and sensitivity to parking error, because the arm must compensate for wherever the base actually stopped.

Whole-body motion treats the base and arm as one kinematic chain, with the base's planar position and heading as extra joints. The planner or controller can then move both at once, for example sweeping the base along a shelf while the arm picks, or using a small base motion to extend reach when the arm is near its limit. Whole-body control resolves the redundancy with priorities: keep the gripper on the target first, avoid obstacles and joint limits second, prefer comfortable arm configurations third. The coupled system is more capable but harder to plan, to verify and to make safe, and wheels add nonholonomic constraints that a simple joint model does not capture.

The error budget

A fixed arm in a calibrated cell can be accurate to a fraction of a millimeter. A mobile manipulator adds the base localization error, which with lidar localization in a known map is typically a few centimeters in position and around a degree in heading, plus any error in the map itself. Heading error matters most at long reach: a one-degree heading error moves the gripper by \( r \sin 1^\circ \approx 0.017 r \), about 15 mm at 0.9 m reach.

Independent error sources combine approximately as the root sum of squares. The total must be smaller than what the grasp tolerates, which lesson 5 estimated from the gripper's clearance and the antipodal margin. If it is not, the system needs a correction: re-perceiving the target after parking, so that the arm moves relative to what the camera sees rather than relative to the map; docking against fixed features; or using fiducial markers on stations.

# Mobile manipulation error budget: root sum of squares versus Monte Carlo, with and without re-perception
import numpy as np
rng = np.random.default_rng(10)
r = 0.9                                       # reach from base middle to gripper, m
sig_xy, sig_th = 0.020, np.radians(1.0)       # base localization (per axis) and heading
sig_arm, sig_cam = 0.0005, 0.003              # arm repeatability and camera pose error (per axis)
tol = 0.010                                   # lateral tolerance the grasp accepts, m
n = 200000
# lateral error at the gripper when moving to a map-based target (no re-perception)
lat_map = rng.normal(0, sig_xy, n) + r * np.sin(rng.normal(0, sig_th, n)) + rng.normal(0, sig_arm, n) + rng.normal(0, sig_cam, n)
# after parking, the wrist camera re-measures the target: base errors cancel, camera and arm remain
lat_rep = rng.normal(0, sig_cam, n) + rng.normal(0, sig_arm, n)
rss = np.sqrt(sig_xy**2 + (r*sig_th)**2 + sig_arm**2 + sig_cam**2)
print(f"RSS lateral sigma (map-based): {1000*rss:.1f} mm")
print(f"Monte Carlo sigma (map-based): {1000*lat_map.std():.1f} mm, within ±10 mm: {100*np.mean(abs(lat_map) < tol):.1f} %")
print(f"Monte Carlo sigma (re-perceived): {1000*lat_rep.std():.1f} mm, within ±10 mm: {100*np.mean(abs(lat_rep) < tol):.2f} %")

RSS lateral sigma (map-based): 25.6 mm
Monte Carlo sigma (map-based): 25.7 mm, within ±10 mm: 30.3 %
Monte Carlo sigma (re-perceived): 3.0 mm, within ±10 mm: 99.90 %

The numbers make the design decision for you. Grasping from the map pose alone would miss a ±10 mm tolerance about seven times in ten, while re-measuring the target with a wrist camera after parking brings success to about 99.9 % in this model. That is why nearly every practical mobile manipulator looks again after it stops, and why camera calibration (ROB-330) is a first-class concern.

Docking, fiducials and station design

Where a mobile manipulator visits the same stations repeatedly, the cheapest accuracy comes from the environment. Mechanical docking guides align the base to a few millimeters. Fiducial markers such as AprilTags or ArUco markers fixed to each station let the robot measure its pose relative to the station with a camera to about a millimeter at short range. Some machine-tending systems touch off against reference features with the gripper, as a CNC machine probes a part. Designing stations with these features is part of deploying mobile manipulators, not an afterthought.

Interacting with the environment

Mobile manipulators in buildings must interact with things designed for people: doors, drawers, elevator buttons, light switches, carts. Opening a door means grasping a handle and moving along a circular path whose radius and hinge location are only roughly known; lesson 11 shows how to estimate them from motion. Force control (ROB-300) is essential here, because a stiff arm pulling along a slightly wrong arc would fight the hinge and either stall or damage something. Compliance, either from impedance control or from a compliant wrist, lets the constraint guide the motion.

Whole-body considerations return in these tasks: a robot opening a heavy door may need to move its base while pulling, and pushing a cart requires the base to supply the force while the arm holds the handle. The base's own stability, the risk of tipping when the arm extends with a load, must be checked like any placement in lesson 8: the combined center of mass must stay inside the wheel support polygon with margin, including during accelerations.

Safety for mobile manipulators

Mobile manipulators combine the hazards of mobile robots, such as collisions with people and objects while driving, with those of arms, such as pinching and impact. Standards address both. ANSI/A3 R15.08, in parts published in 2020 and 2023, covers industrial mobile robots and is used in both the United States and Canada; ISO 3691-4 covers driverless industrial trucks; and the robot standards ISO 10218-1 and -2 (2025 editions), adopted in Canada through CSA Z434 and in the United States through ANSI/A3 R15.06, cover the arm. Collaborative operation and power and force limits from those standards apply when the arm works near people. A typical risk assessment defines safety fields for the base's scanners that shrink as it slows, limits arm speed and force when people are near, and specifies what the robot does when someone enters its workspace while the arm is extended.