The central challenge in robotics is the "curse of dimensionality." As a robot gains more joints—moving from a simple mobile base to a 15-degree-of-freedom (DoF) humanoid—the space of possible configurations grows exponentially. Searching this space for a collision-free path becomes computationally prohibitive for real-time applications.
To solve this, researchers have converged on a powerful intuition: don't solve the whole problem at once. Instead, decouple the high-dimensional system into a hierarchy of simpler, lower-dimensional sub-problems. By solving for the "primary" constraints first and refining the details later, planners can find valid trajectories in milliseconds rather than hours.
The Geometry of Constraints
The earliest versions of this idea focused on the physical structure of the robot. In his foundational work on the PUMA arm, Rodney A. Brooks proposed that a multi-jointed manipulator should be treated as a sequence of segments. In Find-Path for a PUMA-Class Robot, Brooks showed that by propagating constraints from the base segments to the distal ones, a planner can solve for motion in a prioritized sequence.
This hierarchical decomposition works because the motion of the "base" or "primary" links typically dictates the feasibility of the entire path. S. M. Udupa pioneered this in Collision Detection and Avoidance in Computer Controlled Manipulators, arguing that we should route the most constrained components first. If the main arm cannot clear an obstacle, there is no point in calculating the finger positions. By treating secondary links as mere "adjustments" to a primary trajectory, the search space collapses from a high-dimensional volume into a manageable sequence of lower-dimensional searches.
Decoupling Global Strategy from Local Refinement
A second major facet of this direction is the separation of where to go from how to get there. This is often framed as a decoupling of holonomic "rough drafts" and kinematically feasible "refinements."
Jean-Claude Latombe demonstrated this in A Fast Path Planner for a Car-Like Indoor Mobile Robot, where a global path is found ignoring complex steering constraints, which are then layered back in during a local refinement phase. This approach was later extended to robots with complex dynamics by Erion Plaku. In Robot Motion Planning with Dynamics as Hybrid Search, Plaku uses low-dimensional workspace abstractions (like simple 2D triangulations) to guide high-dimensional sampling. The abstraction acts as a "scaffold," ensuring the high-dimensional search doesn't waste time exploring dead ends in the workspace.
Task-Space vs Joint-Space
Modern high-DoF systems, such as the 15+ DoF arm-hand systems studied by Alexander Alexiev and Sangbae Kim, apply this decoupling to the distinction between the "task space" (where the fingertips are) and the "joint space" (the angles of every motor).
In Hierarchical Reactive Grasping via Task-Space Velocity Fields and Joint-Space Quadratic Programming, they show that planning for the end-effectors in a low-dimensional space allows for real-time reactivity. The complex "internal" motion of the arm is handled by a local optimizer (Quadratic Programming) that simply tries to follow the task-space path while avoiding self-collision. This prevents the computational cost from scaling cubically with the number of joints, enabling robots to catch moving objects or react to disturbances instantly.
Sequencing and Motion Primitives
When the environment is cluttered, even the "base" path can be complex. Recent work like TopAY by Long Xu and colleagues refines the hierarchy further. They first identify distinct "topological paths" (e.g., go left of the chair vs. right of the chair) for the mobile base, and only then sample manipulator configurations for those specific routes.
This strategy of sequencing is also vital for cooperative tasks. Atsushi Yamashita showed in Motion Planning for Cooperative Transportation that instead of solving for all degrees of freedom of a multi-robot team simultaneously, one can alternate between types of movement—such as horizontal translation followed by vertical pose adjustment.
Finally, for highly dynamic robots like quadrupeds, Aaron Ames demonstrates in Robust Locomotion on Legged Robots through Planning on Motion Primitive Graphs that the entire joint-state space can be abstracted into a graph of discrete "behaviors." Planning then becomes a search over this graph, allowing the robot to "think" in terms of steps and recoveries rather than individual motor torques.
Why It Works: The Hierarchy of Importance
The success of these methods rests on a single realization: not all degrees of freedom are created equal. In almost every robotic task, a small subset of variables (the base position, the gripper location, or the primary joints) defines the success of the plan. By solving these "bottleneck" variables first, the search space for the remaining degrees of freedom is so severely constrained that the final solution becomes trivial to compute. This hierarchical approach transforms the curse of dimensionality into a series of manageable, low-dimensional problems.