Warm-Starting Non-Convex Trajectory Optimization

10 concepts10 papers1 proposal

Trajectory optimization is a cornerstone of modern robotics, enabling quadrotors to navigate dense forests, humanoids to walk, and robotic arms to catch fast-moving objects. However, generating these trajectories requires solving complex, non-linear, and non-convex optimization problems.

In a real-time deployment, a robot cannot afford to wait seconds for a solver to converge. If you initialize an optimizer from scratch—a "cold start" using random values or zero control inputs—the solver must blindly navigate a mathematical landscape cluttered with local minima and obstacles. This process is slow and frequently gets trapped in poor, dynamically infeasible solutions.

Trajectory warm-starting solves this by providing numerical optimizers with a high-quality initial guess. Instead of starting blindly, the solver begins its search directly inside the "basin of attraction" of a near-optimal solution. By combining global approximation methods with fast local refinement, warm-starting transforms intractable optimization problems into real-time control loops.


The Geometry of the Basin of Attraction

To understand why warm-starting works, consider the optimization landscape as a terrain of hills and valleys, where the lowest point represents the most efficient, collision-free trajectory.

Numerical solvers (such as Sequential Quadratic Programming or Differential Dynamic Programming) are local optimization tools. They are exceptional at finding the bottom of a valley if they are dropped directly into it, but they cannot see over hills to find better valleys. The region of the landscape where a local solver will successfully slide down into a specific minimum is called its basin of attraction.

A cold start often drops the solver on top of a hill or in a completely unrelated valley. A warm start, by contrast, uses prior knowledge to drop the solver directly into a high-quality basin of attraction. From there, the local solver only needs a few iterations to polish the rough guess into a mathematically precise, dynamically feasible trajectory.

Researchers typically generate these high-quality initial guesses through two main paradigms: data-driven learning and structural continuity.


Data-Driven Warm-Starts: Learning the Mapping

One powerful way to find a good basin of attraction is to learn it from experience. By training a machine learning model offline on thousands of successfully optimized trajectories, the model learns a direct mapping from environmental or task parameters to a nominal trajectory.

Regression and Sequence Models

For highly dynamic tasks like robotic catching, Lampariello et al. demonstrated that regression models can map target object characteristics directly to optimal trajectory parameters. This mapping bypasses the need for heuristic starting points, enabling a simplified online optimizer to converge rapidly enough for real-time execution.

Similarly, in humanoid locomotion, Lembono and Mastalli introduced a "Memory of Motion" framework. Their regression model maps high-level task parameters to high-dimensional trajectories, reducing the computational effort required for an optimal control solver to converge by up to 70%.

As robot behaviors become more complex, sequence models offer an elegant way to capture temporal dependencies. Gammelli et al. leveraged a Transformer architecture to model trajectories as sequences. The Transformer generates a near-optimal initial trajectory guess that guides a classical Model Predictive Control (MPC) solver, achieving massive speedups in non-convex environments.

When a robot's state space includes discrete transitions—such as a legged robot deciding whether to trot, bound, or walk—the optimization landscape becomes fragmented into distinct zones. Mastalli et al. addressed this by using a predictive model to determine the discrete gait region first, and then warm-starting the continuous trajectory optimizer with the mean trajectory of that specific region. This targeted initialization provides up to a 60x computational speedup by ensuring the solver begins inside the correct behavioral basin.

[Task Context / Environment] 
         │
         ▼
 ┌───────────────┐
 │ Learned Model │ (e.g., Transformer, Regression, Memory of Motion)
 └───────┬───────┘
         │
         ▼ [Nominal Trajectory Guess] (Inside the correct basin of attraction)
 ┌───────────────┐
 │ Local Solver  │ (Fast numerical refinement)
 └───────┬───────┘
         │
         ▼
 [Optimal, Feasible Trajectory]

Structural Continuity: Exploiting Problem Relationships

We do not always need a machine learning model to predict a warm start. In many robotics applications, the optimization problems solved at step $t$ and step $t+1$ are deeply related. We can exploit this structural continuity across time, space, or objectives.

Temporal and Spatial Continuity

In receding-horizon planning, a robot continuously re-plans its path as it moves. For high-speed quadrotor flight, Ren et al. designed a "Bubble Planner" that extracts safe corridors of overlapping spheres. By reusing waypoints from the overlapping portions of a previously computed plan, the optimizer avoids falling into suboptimal local minima caused by rapid changes in velocity.

This concept also applies to graph-based kinodynamic motion planning, where a robot searches a tree of possible actions. Natarajan et al. showed that when evaluating a new "child" trajectory segment, the optimizer can be warm-started using the optimized solution of its "parent" node. This strategy drastically reduces the time spent on local optimization during the graph search.

Parametric Continuity in Multi-Objective Spaces

Sometimes the continuity exists not in time, but in the problem's objective space. In Multi-Objective Ergodic Search—where a robot must balance conflicting goals like searching for survivors versus locating hazardous leaks—finding the Pareto-optimal front requires solving the optimization problem across various objective weights.

Instead of solving each weight configuration independently, Ren et al. (2022) and Ren et al. (2023) developed a sequential local optimization framework. Because adjacent points in the objective weight space have similar optimal trajectories, the authors reused the optimized trajectory from a neighboring weight vector as the initial guess for the next. This sequential warm-starting approach cuts total optimization iterations by more than half.

Extrapolation and Projection

We can also exploit continuity by looking at the trajectory of the optimization process itself. In graph matching problems, Wang et al. treated the optimization path as a smooth curve, calculating the "velocity" of previous solution points to extrapolate and predict the next starting position.

Even in combinatorial optimization, such as the Traveling Salesman Problem, structural commonalities can be harvested. Braun and Buhmann demonstrated that by finding an "average trajectory" across noisy problem instances, new instances can be projected onto this backbone to establish an initial ordering, which local search operations then rapidly refine.


Why It Works: The Division of Labor

The fundamental reason warm-starting is so successful across these diverse domains lies in a clean division of labor between global approximation and local refinement:

  1. Global Methods (ML, Heuristics, Sequential Continuity): These methods excel at understanding the macro-structure of the problem. They understand the "big picture"—such as which side of an obstacle to pass on, or what gait pattern to use. They are computationally cheap but lack the precision to satisfy exact physical constraints or collision boundaries.
  2. Local Methods (Numerical Optimizers): These methods are blind to the big picture but possess mathematical rigor. Given a rough guess, they can rapidly adjust forces, torques, and positions to satisfy strict physics and safety constraints.

By using a global method to select the basin of attraction (the warm start) and a local method to find the exact minimum, warm-starting combines the speed of intuition with the safety of exact mathematics.

Key Foundational Work

Trajectory planning for optimal robot catching in real-time (2011) — contributed 2 concept elements to this direction.
The Noisy Euclidean Traveling Salesman Problem and Learning (2001) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_trajectory_warm_starting_2ff78717,
  title = {trajectory warm-starting},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/trajectory-warm-starting/rdr_2ff7871791b04825bde49489962c4513},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_trajectory_warm_starting_2ff78717_v1,
  title = {trajectory warm-starting},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/trajectory-warm-starting/rdr_2ff7871791b04825bde49489962c4513/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake