Integrating Global Pathfinding with Local Trajectory Optimization

531 concepts397 papers5 proposals

Imagine driving through an unfamiliar city using only a close-up view of your immediate surroundings. You would easily avoid pedestrians and potholes, but you would inevitably get trapped in dead ends or cul-de-sacs. Conversely, if you only looked at a static city map, you would plan a perfect route but crash into the first dynamic obstacle that crossed your path.

This scenario captures the fundamental tension in autonomous motion planning. Robots must navigate environments that are both globally complex (requiring long-range strategic foresight) and locally dynamic (requiring high-frequency reactivity).

Historically, planners fell into two distinct camps. Global planners (such as $A^*$, Ant Colony Optimization, or Rapidly-exploring Random Trees) search the entire environment to find a path to the goal. While they guarantee reachability and avoid large-scale traps, they are computationally heavy and struggle to adapt to real-time changes. Local planners (such as Model Predictive Control, Timed Elastic Bands, or Artificial Potential Fields) optimize trajectories over a short horizon using immediate sensor data. They are fast and respect complex vehicle dynamics, but they suffer from non-convexity—getting trapped in local minima when obstacles block their greedy paths.

The solution that has emerged across recent research is a hierarchical integration: coupling global search with local optimization to achieve both global reachability and local feasibility in real time.

The Core Mechanism: Hierarchical Decomposition

The insight behind this hybrid approach is to decouple the planning problem into a hierarchical architecture: a global "navigator" and a local "driver." Instead of forcing a single algorithm to solve both long-range topology and short-range physics simultaneously, the system splits the computational burden.

The global layer operates on a simplified representation of the world (such as a static map, a low-fidelity grid, or a sparse graph) to find a coarse, collision-free path. The local layer then treats this global path not as an immutable track, but as a flexible guide. It optimizes the robot's actual control inputs over a short time horizon, deviating from the global path to dodge dynamic obstacles and returning to it once the danger has passed.

This synergy allows robots to navigate highly unstructured or adversarial environments. For instance, Ray in Navigation in Adversarial Environments Guided by PRA* and a Local RL Planner demonstrates that pairing a global pathfinder with a local reinforcement learning agent allows units to proactively dodge dynamic threats while maintaining steady progress toward a long-range destination.

How the Layers Communicate

The critical engineering challenge in these hybrid systems is the interface: how does the global path guide the local optimizer without overwhelming it? Researchers utilize three primary coupling mechanisms.

1. Global Paths as Directional Constraints and Corridors

One approach is to extract key topological features from the global path—such as turning points or bounding corridors—and inject them directly into the local planner's objective function.

In Global and local path planning of robots combining ACO and dynamic window algorithm, Lu integrates global path turning points as directional constraints within a local Dynamic Window Approach (DWA). By adding a heading evaluation term that penalizes deviations from the pre-calculated global route, the local planner is prevented from losing track of the long-term goal during obstacle avoidance.

Similarly, Zhou et al. in Robust and Efficient Quadrotor Trajectory Generation for Fast Autonomous Flight use a kinodynamic global search to define a safe corridor, which then constrains a gradient-based B-spline local optimizer. This ensures that high-speed quadrotor flight remains globally safe and locally smooth.

2. Global Guidance Features for Learning Policies

For learning-based local planners, the global path can be fed directly into the model as an input feature or a dense reward signal. This resolves the notorious "sparse reward" problem in reinforcement learning, where an agent wandering randomly in a large map rarely reaches the goal.

The authors of Mobile Robot Path Planning in Dynamic Environments through Globally Guided Reinforcement Learning introduce a framework where a pre-computed global path acts as a guidance channel for a reinforcement learning agent. The agent treats the global route as a flexible hint, allowing it to execute reactive local maneuvers around dynamic obstacles while retaining the global context needed to prevent it from getting stuck.

3. Two-Stage Optimization and Model Abstraction

Another powerful paradigm uses a simplified model for the global stage to find a high-quality initial guess, which is then refined by a high-fidelity local non-linear optimizer.

Because motion planning in cluttered environments is highly non-convex, non-linear solvers are prone to failing or finding poor local minima if initialized poorly. Eiras in A Two-Stage Optimization-Based Motion Planner for Safe Urban Driving solves this by using Mixed-Integer Linear Programming (MILP) on a simplified model to identify a "safe skeleton" subspace. This linear solution provides a high-quality initialization for a high-fidelity non-linear programming (NLP) solver. By using the global layer to isolate the correct convex valley, the local solver converges rapidly to a safe, dynamically feasible trajectory.

A similar abstraction strategy is used by Pěnička et al. in Minimum-Time Quadrotor Waypoint Flight in Cluttered Environments, where a simplified point-mass model guides a full-dynamics simulation, dramatically reducing the search space for aggressive quadrotor flight.

Why It Works: Breaking the Curse of Non-Convexity

The fundamental reason this hybrid approach works is that it breaks the curse of dimensionality and non-convexity simultaneously.

If you try to plan a trajectory from start to finish using full robot dynamics and all obstacles simultaneously, the optimization problem is too complex to solve in real time. By separating the problem, the global planner handles the combinatorial complexity of the environment (e.g., deciding whether to go left or right around an obstacle) using a low-dimensional model. The local planner handles the differential complexity of the robot (e.g., actuator limits, friction, and immediate safety boundaries) over a short time horizon.

As demonstrated by Pallar et al. in Optimal Trajectory Planning for Cooperative Manipulation with Multiple Quadrotors Using Control Barrier Functions, integrating global pathfinding with local safety-constrained optimization prevents local safety mechanisms from becoming trapped in local minima. The global path acts as an anchor, pulling the local optimizer through complex obstacle fields that would otherwise paralyze a purely reactive system.

History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_hybrid_path_planning_862c1571,
  title = {hybrid path planning},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/hybrid-path-planning/rdr_862c1571c6d24a429bcee13a46886eed},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_hybrid_path_planning_862c1571_v1,
  title = {hybrid path planning},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/hybrid-path-planning/rdr_862c1571c6d24a429bcee13a46886eed/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake