Bridging Direct and Indirect Methods via Sequential Convex Programming

This direction is merged into another direction: generalization to nonlinear systems. Its last version stays readable and citable here.

Version 1 of 19 papers19 conceptsfrozen 2026-10-04

In the world of robotic motion planning and spacecraft navigation, solving optimal control problems is a balancing act between reliability and precision. Engineers typically choose between two worlds: direct methods, which are robust but computationally heavy, and indirect methods, which are lightning-fast but notoriously difficult to initialize.

Recent research led by Riccardo Bonalli, Marco Pavone, and their colleagues has built a mathematical bridge between these two worlds. By using Sequential Convex Programming (SCP) not just as a solver, but as a "translator" for hidden physical variables, they have unlocked a way to achieve both global robustness and local high-speed convergence.

The Dilemma: Search vs. Precision

To understand why this bridge is necessary, we must look at how these two methods "see" a problem:

  1. Direct Methods (SCP): These methods discretize the trajectory and solve a sequence of easier, convex sub-problems. They have a wide "convergence basin," meaning they can find a solution even from a poor initial guess. However, they often struggle to reach high-precision optimality quickly.
  2. Indirect Methods (Shooting): These treat the problem as a Boundary Value Problem based on the Pontryagin Maximum Principle (PMP). If you have the correct "costates" (mathematical shadows of the system's constraints), these methods converge quadratically—extremely fast. The catch? If your initial guess for the costates is even slightly off, the method usually fails entirely.

The fundamental problem is that costates (or adjoint variables) are non-intuitive. While a human can guess a rough path for a robot (the state), guessing the sensitivity of the cost to every constraint (the costate) is nearly impossible.

The Bridge: Multipliers as Costates

The breakthrough described in Analysis of Theoretical and Numerical Properties of Sequential Convex Programming for Continuous-Time Optimal Control lies in the realization that the "waste products" of direct methods are actually the "fuel" for indirect methods.

When SCP solves a convex sub-problem, it generates Lagrange multipliers. Riccardo Bonalli and Thomas Lew proved that as SCP converges, these discrete Lagrange multipliers converge to the continuous-time costates of the original non-convex problem.

This means that SCP is doing two things at once:

  • It is finding the optimal path (the Direct task).
  • It is implicitly calculating the exact costate information needed for the PMP (the Indirect task).

A Two-Stage Strategy for Speed

This theoretical link enables a powerful two-stage optimization framework, exemplified by the GuSTO (Guaranteed Sequential Trajectory Optimization) algorithm. In GuSTO: Guaranteed Sequential Trajectory optimization via Sequential Convex Programming, Abhishek Cauligi, Andrew Bylard, and the team demonstrate a hybrid workflow:

  1. Stage 1 (The Scout): Run SCP for a few iterations. Because SCP is robust, it quickly moves the trajectory into the right neighborhood, even in cluttered environments with complex manifold constraints.
  2. Stage 2 (The Sprinter): Extract the Lagrange multipliers from the SCP solver and use them to "warm-start" an indirect shooting method.

Because the multipliers provide a high-quality estimate of the costates, the indirect method starts within its narrow convergence window. It then "sprints" to the final, high-precision solution with quadratic speed.

Why It Works: Theoretical Rigor

This isn't just a heuristic trick. In Analysis of Theoretical and Numerical Properties of Sequential Convex Programming for Continuous-Time Optimal Control, the authors provide the first unifying analysis proving that SCP solutions in continuous time satisfy the Pontryagin Maximum Principle.

This theoretical foundation ensures that the "bridge" is stable. By mapping dual variables from convex sub-problems to the adjoint variables of the non-convex whole, researchers have turned a mathematical abstraction into a practical tool for real-time robotics. In high-dimensional tasks like spacecraft navigation, this hybrid approach can reduce computation time by over 70% compared to using direct methods alone, all while maintaining the mathematical guarantees required for safety-critical missions.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_bridging_direct_and_indirect_methods_via_2c6b8dd4,
  title = {Bridging Direct and Indirect Methods via Sequential Convex Programming},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/bridging-direct-and-indirect-methods-via-sequential-convex-programming/rdr_2c6b8dd4693f4a89ba1ad72064df0341},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_bridging_direct_and_indirect_methods_via_2c6b8dd4_v1,
  title = {Bridging Direct and Indirect Methods via Sequential Convex Programming},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/bridging-direct-and-indirect-methods-via-sequential-convex-programming/rdr_2c6b8dd4693f4a89ba1ad72064df0341/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}