Depth & heading control for AUVs

The mission
The vehicle must reach a commanded depth and heading while respecting the limits of its actuators. The side view shows depth and pitch; the top view shows heading. The study considers a nonlinear underwater vehicle controlled using reduced, linearized prediction models.

Why model mismatch matters
A linearized model makes repeated optimization practical, but it does not capture every nonlinear effect or coupling. In the studied baseline, this mismatch produces persistent depth and heading errors. This is a limitation of this baseline formulation, not of every possible predictive controller.

Give each layer a clear job
The adaptive reference layer adjusts the depth and heading references supplied to model predictive control. The predictive controller computes control inputs while incorporating state, actuator, and rate limits. Allocation converts the virtual control channels into commands for the physical control surfaces.

Make actuator limits visible
Each actuator has a permitted range and a limit on how quickly its command can change. The four tail planes combine vertical and horizontal control demands. The optimization accounts for their allocated commands, so a feasible virtual input must also respect the physical actuator bounds.

Designing MPC: why an infinite horizon?
Conventional model predictive control usually optimizes over a finite horizon. A short prediction window may end before the vehicle reaches its target. Here, an infinite horizon represents the full approach from the current state to the commanded final target, including settling.

Map time; the problem is still continuous
Following the infinite-horizon domain-transformation approach of Fahroo and Ross, we map physical time onto the interval from zero to one, excluding one. In our normalization, time equals tau divided by one minus tau. The transformed optimal control problem is still continuous in time; this mapping is not yet discretization.

Bernstein transcription: from OCP to NLP
Next, we choose approximation degree N and represent states and inputs with Bernstein polynomials. Their coefficients become optimization variables. Transcribing the cost, dynamics, and constraints produces a finite nonlinear program. Solving it determines the coefficients, and therefore the represented state and input trajectories.

Convex-hull containment at every degree
Bernstein basis functions are nonnegative and sum to one. The curve therefore stays inside the convex hull of its control points, even at a small degree N. Coefficient bounds in a convex admissible set bound the represented curve. Approximation accuracy remains a separate question.

Replan from new measurements
At each update, the controller uses the latest measured state to solve a new constrained prediction problem. It applies the current control input, receives fresh measurements, and repeats. The adaptive reference layer adjusts the targets used by this same predictive control structure.

Compare the same controller with and without augmentation
The manuscript compares predictive control alone with the same controller assisted by adaptive reference management. Model, costs, allocation, and constraints are retained to isolate the added layer. The reported simulations reduce steady-state depth and heading errors while respecting actuator magnitude and rate limits in the tested cases.

Interpret the evidence
The reported study uses a nonlinear simulation plant at two forward speeds, with calm-water test conditions. These animations illustrate the architecture and constraint geometry. They are not a replay of the simulation data, and they do not establish performance in arbitrary wave conditions.