Control and allocation¶

Marine vehicles are usually controlled in a cascade: guidance makes a desired velocity or pose, a controller turns it into a desired force and moment (a wrench), and thrust allocation distributes the wrench over the actuators.

flowchart TB
  MIS["Mission<br/>waypoints, behaviors"] --> GUID["Guidance<br/>LOS, path following"]
  GUID -->|"desired velocity / heading"| CTRL["Motion controller<br/>PID, sliding mode, MPC"]
  CTRL -->|"wrench τ"| ALLOC["Thrust allocation"]
  ALLOC -->|"thrust per actuator"| MAP["Thrust → command map<br/>(thruster curve)"]
  MAP -->|"PWM, RPM, or CAN"| ESC["ESCs / motor drivers"]
  EST["State estimate<br/>(odometry)"] --> GUID
  EST --> CTRL

Typical ROS 2 messages along the chain: nav_msgs/Path or waypoints → geometry_msgs/Twist → geometry_msgs/Wrench → per-thruster std_msgs/Float64 or a ros2_control command interface.

Packages¶

Project	What it does	ROS	Vehicles	Activity
auv_controllers Robotic Decision Making Lab (Oregon State University)	ros2_control controllers for underwater vehicles, including twist, trajectory, impedance, and whole-body controllers, a thruster allocation matrix controller, and thruster controllers. Supports ROS 2 Rolling and Jazzy.	ROS 2	ROV AUV	Active Aug 2026
mvp_control URI Smart Ocean Systems Lab	Optimization-based control allocation for multi-thruster vehicles, including articulated thrusters. ROS 2 on jazzy-devel, ROS 1 on noetic-devel.	ROS 2 ROS 1	USV AUV ROV	Active Sep 2026
mvp_mission URI Smart Ocean Systems Lab	Behavior-based mission and guidance layer for MVP vehicles. ROS 2 on jazzy-devel.	ROS 2 ROS 1	USV AUV ROV	Active Sep 2026
thruster_manager Centrale Nantes	Maps a desired wrench to thruster forces. Reads the thruster layout from the URDF on robot_description; usable as a node, a component, or a C++ class (Eigen).	ROS 2	ROV AUV	Low activity Apr 2025
ros2_control ros-controls	Standard ROS 2 framework for hardware interfaces and real-time controllers. Underwater controllers such as auv_controllers build on it.	ROS 2	Any	Active Oct 2026

Thrust allocation¶

Given a desired wrench \( \tau \) (forces and moments in the body frame) and thrust vector \( f \) for \( n \) thrusters, the allocation matrix \( T \) maps thrust to wrench:

\[ \tau = T(\alpha)\, f, \qquad T \in \mathbb{R}^{6 \times n} \]

Column \( i \) of \( T \) is \( [\,d_i ;\; r_i \times d_i\,] \), where \( d_i \) is the unit thrust direction and \( r_i \) the thruster position relative to base_link. For azimuth thrusters, \( T \) also depends on the angles \( \alpha \). Read the positions and directions from the URDF so that the controller and the simulation use the same geometry (thruster_manager does this).

Methods¶

Method	Handles thrust limits	Azimuth thrusters	Cost	Use when
Pseudo-inverse \( f = T^{+}\tau \)	No	No (fixed angles)	One matrix product	Fixed thrusters, and demands stay well inside the limits
Weighted pseudo-inverse	No, but can favor some thrusters	No	One matrix product	Some thrusters are weaker, less efficient, or noisier
Pseudo-inverse + uniform scaling	Keeps wrench direction, shrinks magnitude	No	Cheap	Simple vehicles that saturate sometimes
Quadratic programming (QP)	Yes: hard limits, rate limits, forbidden zones	Yes	Solver per cycle	Real vehicles that saturate, azimuth thrusters, prioritized axes

The pseudo-inverse is valid for non-square \( T \) (more thrusters than controlled axes): it gives the minimum-norm thrust. Its problem is that it ignores limits, not that \( T \) is non-square.

Tested example¶

A vectored layout with four horizontal thrusters, controlling surge, sway, and yaw:

allocation_example.py (output below is real)

import numpy as np

# Thruster x, y (m, FLU body frame) and thrust direction angle (rad, from +x, CCW)
thrusters = np.array([
    [ 0.15,  0.11, np.deg2rad(-45)],   # front left
    [ 0.15, -0.11, np.deg2rad( 45)],   # front right
    [-0.15,  0.11, np.deg2rad(-135)],  # rear left
    [-0.15, -0.11, np.deg2rad( 135)],  # rear right
])
x, y, a = thrusters.T

# Rows: surge force, sway force, yaw moment (z component of r x d)
T = np.vstack([np.cos(a), np.sin(a), x * np.sin(a) - y * np.cos(a)])
T_pinv = np.linalg.pinv(T)

tau = np.array([30.0, 0.0, 3.0])       # 30 N forward, 3 N*m yaw
f = T_pinv @ tau                        # [  6.53  14.69  -6.53 -14.69]

f_max = 10.0                            # thruster limit (N)
f_clip = np.clip(f, -f_max, f_max)      # wrench -> [23.4, 0, 1.28]  direction changes
f_scaled = f * min(1.0, f_max / np.abs(f).max())  # wrench -> [20.4, 0, 2.04]  direction kept

Clipping each thruster on its own changes the ratio between surge and yaw (here from 10:1 to 18:1), so the vehicle turns less than commanded. Scaling keeps the ratio. A QP allocator can do better: for example, it can keep yaw and give up some surge.

Known issues¶

Pseudo-inverse allocators (thruster_manager, thruster allocation matrix controllers)

Saturation is not handled. Commands beyond the thruster limits are clipped downstream, which changes the wrench direction (see the example). Use uniform scaling or a QP allocator.
Ill-conditioned layouts. If thrusters are nearly parallel, \( T \) is close to singular and small wrench changes give large thrust changes. Check np.linalg.cond(T) at design time.
Dead zones. Many thrusters give no thrust below a minimum command. Small demands then do nothing, which can cause limit cycles in station keeping. Compensate in the thrust-to-command map.

QP allocators (mvp_control and others)

Solve time varies. Monitor the solve time; reduce the problem size or the control rate if it causes jitter.
More parameters. Weights and constraints need tuning. Start with equal weights and add constraints one at a time.

All allocators

The thruster curve matters as much as the allocation. Thrust is roughly proportional to RPM squared and is asymmetric forward and reverse. Use the manufacturer's thrust curve at your supply voltage, or measure it on a test stand.
Thrust falls with battery voltage. Compensate, or retune when the battery type changes.
Thruster interaction. Thrusters close to each other or to the hull lose thrust. Field results are often lower than test-stand results.

Motion control¶

Station keeping (dynamic positioning)¶

Holds position and heading for inspection, sampling, or intervention. Common approaches:

PID per axis with feed-forward of buoyancy and known currents. Simple and effective for small ROVs. Use anti-windup, because thrusters saturate.
Sliding mode control, robust to model uncertainty and common in the marine control literature.
Impedance and whole-body control for vehicles with manipulators; auv_controllers includes both, together with twist and trajectory controllers.
Model predictive control (MPC) handles constraints explicitly but needs a model and more compute.

Path following and guidance¶

Line-of-sight (LOS) guidance is the classic marine method. For a straight path segment with path angle \( \pi_p \) and cross-track error \( e \), the desired course is

\[ \chi_d = \pi_p - \arctan\!\left(\frac{e}{\Delta}\right) \]

where the lookahead distance \( \Delta \) sets how fast the vehicle converges: a smaller \( \Delta \) converges faster but can oscillate. In a current, add integral action (integral LOS) or estimate the crab angle, because heading and course differ. The sign depends on how you define \( e \); keep it consistent with your frame convention.

Other options: pure pursuit, vector field guidance, and Nav2 controllers on surface vehicles (for example Regulated Pure Pursuit), with the vehicle modeled as differential drive.

Underactuated vehicles¶

Torpedo AUVs and most ASVs cannot push sideways. Control surge speed, heading or course, and depth or pitch, and let guidance handle the path. Do not command sway velocity to an underactuated vehicle.

Low-level integration¶

Approach	How	Good for
ros2_control	Hardware interface plugin that writes thruster commands; controllers such as auv_controllers run in the controller manager	Clean real-time separation, standard tooling
Autopilot	ArduSub, ArduPilot Rover, or PX4 runs stabilization; ROS sends setpoints through mavros or DDS	Fast start on BlueROV2 or BlueBoat, built-in failsafes
Direct node	A node converts thrust to PWM through a PWM board (for example PCA9685) or a CAN bus	Small custom vehicles

Keep a hardware failsafe

Whatever the stack, make sure that the thrusters stop when commands stop: use a command timeout in the driver or ESC, and a kill switch the operator can reach. Software watchdogs are not a substitute for a physical cut-off on a vehicle with exposed propellers.