When we visualize a robot navigating a room or moving a robotic arm, we naturally imagine a line connecting point A to point B while avoiding obstacles. Classic path-planning algorithms—such as A*, Rapidly-exploring Random Trees (RRT), or grid-based Q-learning—are highly effective at finding these geometric paths.
However, when these paths are sent directly to physical motors, a fundamental problem emerges: the shortest geometric path is often full of sharp, jagged turns. While a mathematical point can change direction instantaneously, a physical robot possesses mass, momentum, and strict actuator limits. This disconnect between geometric optimization and physical reality has driven a major research direction focused on trajectory smoothness.
The Physical Cost of Geometric Discontinuity
To understand why trajectory smoothness is essential, we must look at what happens to hardware when it attempts to execute a jagged path.
A sharp turn in a grid or a sampling tree represents a point of infinite acceleration change. As DUAN Shuyong et al. observe in their work on Improved A-star Algorithm for Safety Insured Optimal Path with Smoothed Corner Turns, these sudden transitions cause massive spikes in motor current and accelerate mechanical fatigue.
Beyond hardware wear, non-smooth trajectories introduce severe operational inefficiencies:
- Energy Consumption: Frequent changes in direction force the robot to constantly decelerate and accelerate. Carlo Tiseo et al. demonstrate in A Passive Navigation Planning Algorithm for Collision-free Control of Mobile Robots that ignoring high-order smoothness yields high peak momentum, leading to excessive battery drain.
- Mechanical Wear: As highlighted by Ying Li 0004 et al. in A novel Q-learning algorithm based on improved whale optimization algorithm for path planning, algorithms that minimize distance alone create jerky movements that physical hardware struggles to execute, resulting in loose joints and broken actuators over time.
- Environmental Safety: In delicate applications, such as agricultural robotics, mechanical jerk causes vibrations. Zhonghua Miao et al. note in A Fast Path-Planning Method for Continuous Harvesting of Table-Top Grown Strawberries that high trajectory smoothness is mandatory to prevent vibrations from damaging fragile stems or dislodging fruit.
The Mathematical Gap: Geometric vs. Kinematic Continuity
Why do standard planners fail to produce smooth paths? The issue lies in the mathematical definition of continuity along the path.
Most basic planners generate paths with $C^0$ continuity—the positions are connected, but the derivatives (velocity, acceleration, jerk) are not. For example, if a drone flies along a piecewise-linear path generated by a standard sampling algorithm, it would have to stop completely at every vertex to change direction without drifting. As Hongliang Dai explains in EB-RRT* based Navigation Algorithm for UAV, these paths require instantaneous changes in direction that exceed the physical aerodynamic and kinematic limits of the vehicle.
Even when velocity is continuous ($C^1$), higher-order derivatives can still be disjointed. Yong Li et al. explore this in TOPP-DWR: Time-Optimal Path Parameterization of Differential-Driven Wheeled Robots Considering Piecewise-Constant Angular Velocity Constraints, noting that while piecewise-constant angular velocity constraints simplify optimization, the resulting discontinuity in the derivative of velocity introduces vibrations and reduces precision in high-speed maneuvers.
Algorithmic Solutions: Injecting Smoothness into Planning
To bridge the gap between geometry and physics, researchers employ three primary strategies to bake smoothness directly into the planning pipeline.
1. Smoothness-Aware Search Heuristics
Instead of generating a jagged path and trying to smooth it afterward, planners can penalize sharp turns during the search phase.
Wei Chen et al. demonstrate this in A real-time multi-constraints obstacle avoidance method using LiDAR by integrating a corner heuristic directly into the pathfinding transition probabilities. This ensures that the algorithm naturally favors straight lines or gentle curves over stair-step grid patterns. Similarly, Xiaolu Ma addresses this in Mobile Robot Global Path Planning Based on Improved Ant Colony System Algorithm with Potential Field by combining metaheuristic search with potential fields to minimize the frequency of turning points.
2. Continuous Parametric Curves
Another prominent approach involves fitting discrete waypoints with continuous parametric curves, such as Bézier curves, B-splines, or Lamé curves. These functions possess inherent higher-order differentiability, ensuring that velocity and acceleration change smoothly.
Jagged Waypoints: o------>o------>o
\
\-->o
Smoothed Curve: o~~~~~~~~~~\~~~~~~~o
\_____/ (Continuous curvature via B-Splines)
For instance, Yun Du et al. combine swarm intelligence with third-order Bézier curves in Path planning and smoothing of mobile robot based on improved artificial fish swarm algorithm to satisfy steering constraints. For robotic manipulators, Shoujun Zhang utilizes cubic B-spline smoothing in Multi-Strategy Improved Rapid Random Expansion Tree (RRT) Algorithm for Robotic Arm Path Planning, eliminating the jagged transitions typical of random sampling trees and reducing mechanical wear.
3. Trajectory-Level Primitives
Rather than building a path state-by-state, some modern frameworks plan directly in the space of continuous trajectories. By utilizing predefined mathematical building blocks called "movement primitives," the planner is physically incapable of generating an un-smooth action.
As shown by Fabian Otto et al. in Deep Black-Box Reinforcement Learning with Movement Primitives, planning at the trajectory level using these primitives inherently produces human-like, energy-efficient motion. This eliminates the high-frequency jitter common in step-based reinforcement learning, making the resulting policies safe for direct deployment on real-world hardware.