Bridging Global Planning and Local Reactive Control

This direction is split into other directions: global-local planning trade-off. Its last version stays readable and citable here.

Version 1 of 136 papers37 conceptsfrozen 2026-10-04

In robotic navigation, there is a fundamental tension between seeing the forest and seeing the trees. A robot needs a global planner to find a path through a complex maze, avoiding dead ends and concave obstacles that are invisible to local sensors. However, it also needs a local reactive controller to dodge a pedestrian who suddenly steps into its path or to adjust for a physical slip.

Historically, these two systems lived in separate worlds. Global planners like A or RRT are computationally expensive and operate on static maps, while local controllers like the Dynamic Window Approach (DWA) or Artificial Potential Fields are fast but "short-sighted," often getting trapped in local minima. Modern research focuses on fusing these two paradigms into a single, seamless architecture that is both strategically sound and tactically agile.

The Hybrid Fusion Strategy

The most direct way to bridge this gap is to use a global planner to generate a "skeleton" path and a local planner to "flesh it out" in real-time. Kaiyu Li and Xiugang Gong demonstrate this by combining A-Star with the Dynamic Window Approach. In their framework, A* provides the optimal sequence of sub-goals, while DWA ensures the robot's physical constraints (velocity and acceleration) are respected as it avoids dynamic obstacles.

A similar intuition drives the RRT Guided Model Predictive Path Integral (MPPI) method. Here, the Rapid-exploring Random Tree (RRT) acts as a scout, finding a feasible path through a complex environment. This path then "seeds" the MPPI controller—a stochastic optimizer—which refines the trajectory into smooth, high-speed motion. By using the global search to guide the local optimization, the system avoids the "dead ends" that typically plague purely local methods.

Policy Blending and Geometric Consistency

As robots move into more complex configurations, like multi-segment continuum robots or high-DOF manipulators, simple path-following isn't enough. Researchers have turned to more sophisticated mathematical frameworks to blend global intent with local safety.

Nathan D. Ratliff and colleagues introduced Riemannian Motion Policies (RMPs), which treat different robot behaviors (like "reach the goal" or "avoid the elbow collision") as acceleration fields. By assigning a Riemannian metric to each policy, the system can mathematically sum these behaviors into a single, consistent controller. This allows a robot to maintain its global objective while reacting to hundreds of local constraints simultaneously.

For even more dynamic environments, An Thai Le and Jan Peters proposed HiPBOT, which uses Unbalanced Optimal Transport (UOT) to blend policies. Instead of hard-switching between a "global" mode and a "safety" mode, HiPBOT treats policy blending as a convex optimization problem. This ensures that the robot's transition between following a plan and avoiding a collision is smooth and mathematically optimal, preventing the "jitter" often seen in hybrid systems.

Scaling to Unknown and Large-Scale Environments

The challenge intensifies when the environment is not just dynamic, but unknown. In these cases, the "global" map must be built on the fly. Botao He and Ji Zhang addressed this with Air-FAR, a system for aerial navigation. They use a hierarchical visibility graph that updates in milliseconds, allowing a drone to "remember" the global structure of a building it just flew through while using a local planner to navigate tight corridors.

In multi-agent settings, the "local" environment includes the intentions of other robots. Tianze Yang and Guillaume Sartoretti developed ALPHA, which uses a Graph Transformer to combine precise local observations with a "fuzzy" global graph. This allows agents to coordinate high-level strategies (like who goes through a door first) without losing the reactive speed needed to avoid bumping into each other.

Why This Works: The Best of Both Worlds

The success of these hybrid approaches stems from a shared realization: global planners are great at topology, while local planners are great at physics.

  1. Topology: Global planners (like those used by Bernard Faverjon and Pierre Tournassoud) understand the connectivity of space. They know that to get to the other side of a wall, you must first move away from the goal toward a doorway.
  2. Physics: Local planners (like the Deep Reactive Policy by Jason Jingzhou Liu) understand the nuances of momentum, friction, and sensor noise.

By coupling them—whether through A* and DWA fusion, RRT-guided optimization, or hierarchical policy blending—researchers create systems that possess both the foresight to navigate a complex world and the reflexes to survive in it.

Papers in this version

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_bridging_global_planning_and_local_react_0551851a,
  title = {Bridging Global Planning and Local Reactive Control},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/bridging-global-planning-and-local-reactive-control/rdr_0551851a714d4794ab2ae8af86157868},
  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_bridging_global_planning_and_local_react_0551851a_v1,
  title = {Bridging Global Planning and Local Reactive Control},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/bridging-global-planning-and-local-reactive-control/rdr_0551851a714d4794ab2ae8af86157868/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}