Optimal Control and Trajectory Planning
optimal control, trajectory planning, LQR, MPC, cost
1 Application Snapshot
Many real systems do not ask only:
can I stabilize this system?
They ask:
- can I reach the target smoothly?
- can I do it without wasting too much energy?
- can I avoid overshoot, actuator stress, or unsafe states?
- can I plan a whole path, not just react locally?
That is the shift from bare stabilization to optimal control and trajectory planning.
This page is the shortest bridge from feedback language into the idea that a controller is often chosen by optimizing a cost over time.
2 Problem Setting
Suppose the system evolves as
\[ x_{t+1} = f(x_t, u_t) \]
or, in continuous time,
\[ \dot{x}(t) = f(x(t), u(t)). \]
A controller is no longer judged only by whether it keeps the system stable.
It is also judged by a performance objective, such as
\[ \text{cost} = \sum_{t=0}^{T-1} \ell(x_t, u_t) + \phi(x_T) \]
or, in continuous time,
\[ \text{cost} = \int_0^T \ell(x(t), u(t))\,dt + \phi(x(T)). \]
Here:
- \(\ell\) is a running cost
- \(\phi\) is a terminal cost
- \(T\) is a horizon
The planner or controller must now choose actions that respect both:
- the system dynamics
- the cost tradeoffs
3 Why This Math Appears
This language reuses several math layers already on the site:
Control and Dynamics: feedback stabilizes the system, but does not by itself specify the best tradeoff among behaviorsOptimization: objectives, constraints, and tradeoffs become part of the control design itselfStochastic Control and Dynamic Programming: finite-horizon and infinite-horizon planning are sequential optimization problemsNumerical Methods: many trajectory-planning problems are solved approximately, repeatedly, and under discretizationODEs and Dynamical Systems: the state evolution is still the load-bearing object underneath the optimization
So optimal control is not “control plus a bonus objective.” It is the point where dynamics and optimization become one problem.
4 Math Objects In Use
- state \(x_t\)
- control input \(u_t\)
- dynamics law \(f\)
- running cost \(\ell(x_t,u_t)\)
- terminal cost \(\phi(x_T)\)
- horizon \(T\)
- reference trajectory or target set
- sometimes constraints on state or input
In the simplest linear-quadratic setting, these objects become
\[ x_{t+1} = Ax_t + Bu_t \]
and
\[ \ell(x_t,u_t) = x_t^\top Q x_t + u_t^\top R u_t. \]
That is the first clean place where readers usually meet LQR.
5 A Small Worked Walkthrough
Imagine a drone that must move from one altitude to another.
A stabilizing controller could simply try to make the altitude error go to zero.
But several trajectories might all be stable while behaving very differently:
- one reaches the target fast but uses too much thrust
- one uses little thrust but takes too long
- one overshoots and oscillates before settling
Now define a cost like
\[ \sum_{t=0}^{T-1} \left( q(h_t-h_{\mathrm{ref}})^2 + r u_t^2 \right). \]
This cost says:
- penalize being far from the target altitude
- also penalize aggressive thrust commands
The controller now has to balance two goals:
- accuracy
- effort
That tradeoff is the systems-level meaning of optimal control.
Trajectory planning adds one more layer:
- the target may not be a single equilibrium
- it may be a whole path through state space
Then the question becomes not only “how do I stabilize?” but “which state sequence should I follow, and how should I generate the controls that realize it?”
6 Implementation or Computation Note
Three practical branches appear quickly:
Quadratic regulationWhen the model is linear and the cost is quadratic, LQR is the clean first-pass answer.Constraints and receding horizonsWhen state or input limits matter, planning often becomes MPC-like.Uncertainty and partial observabilityWhen the state is estimated rather than measured directly, planning and estimation must interact.
Use these pages as the strongest follow-on support:
7 Failure Modes
- thinking every stable controller is equally good
- writing a cost that ignores actuator effort or physical limits
- confusing trajectory planning with point stabilization
- assuming an optimal open-loop plan remains good after disturbances appear
- forgetting that many “planning” problems are actually solve-estimate-replan loops
8 Paper Bridge
- 16.323 / Principles of Optimal Control -
First pass- official MIT anchor for cost-based control design. Checked2026-04-25. - AA203 / Optimal and Learning-Based Control -
Paper bridge- useful once planning, optimization, and modern control start to merge. Checked2026-04-25.
9 Sources and Further Reading
- 16.323 / Principles of Optimal Control -
First pass- official MIT lecture-note hub for optimal control. Checked2026-04-25. - 16.323 syllabus -
First pass- useful compact framing of what optimal control is trying to solve. Checked2026-04-25. - Topic 18: Linear Quadratic Regulator -
Second pass- a clean official MIT entry into quadratic cost tradeoffs. Checked2026-04-25. - AA203 / Optimal and Learning-Based Control -
Second pass- official Stanford course anchor for planning and optimal control. Checked2026-04-25. - EE364b course information -
Bridge to constrained planning- official Stanford convex-optimization course info with strong optimization-to-control relevance. Checked2026-04-25. - EE364b lectures -
Bridge to constrained planning- useful once you want optimization-backed receding-horizon and constrained-control context. Checked2026-04-25.