Learning a dynamics correction for MPC

A different research question
Adaptive support can compensate for model mismatch while the vehicle operates. The next question is whether the system can learn a reusable description of that mismatch. The objective is to improve prediction and reduce reliance on continuous adaptive compensation, rather than claim an unmeasured energy saving.

The source of the prediction error
Model predictive control uses linearized dynamics to predict the vehicle response. Measurements reveal effects that this nominal model misses. The gap between predicted and observed dynamics supplies information for learning. The network aims to describe that residual, rather than replace the entire controller.

Operate with support while collecting experience
Initially, adaptive reference management and predictive control operate together. As the underwater vehicle follows commands, measurements and applied inputs provide learning data. The adaptive layer remains active during this stage, helping control the vehicle while the network learns from its experience.

Learn a correction, not a second control command
The learning network estimates a correction to the nominal dynamics. That correction is intended for use inside the predictive model. It is not a separate actuator command added to the output. The scene shows the intended reduction of model mismatch, not a measured training curve.

Check the correction before switching
Learning alone is not enough to justify a handover. The correction must be checked on relevant conditions, and its effect on tracking and constraints must be assessed. Adaptive support stays available until those checks justify the transition. A training loss by itself is not proof of closed-loop reliability.

The intended handover
After validation, the intended transition disables the adaptive reference contribution. The learned dynamics correction remains inside model predictive control. The optimization structure and actuator limits are retained. This is the research handover being investigated, not a claim that the network has learned every possible condition.

A new command after handover
A new depth and heading command tests the post-handover configuration. Predictive control uses the nominal model plus the learned correction. Tracking, constraint compliance, and sensitivity to unfamiliar conditions must all be evaluated. The motion here illustrates the intended behavior rather than a validated experiment.

What success would mean
The expected value is retained model knowledge and less dependence on continuous adaptive correction within validated conditions. Whether this also saves computation or electrical energy requires measurement: network inference and optimization have costs of their own. This project remains a research concept under development.