Most classical pathfinding algorithms—from A* to RRT—treat a robot's journey as a sequence of discrete waypoints. While this abstraction simplifies the search for a collision-free route, it creates a fundamental mismatch with the physical world. A robot is not a teleporter; it has mass, inertia, and actuators that cannot instantly change velocity. A jagged, piecewise-linear path requires infinite acceleration at every corner, leading to jerky movements, mechanical wear, and tracking errors.
To bridge this gap, researchers have converged on a shared strategy: representing trajectories not as sets of points, but as continuous, differentiable curves—primarily splines. By shifting the optimization problem from the space of coordinates to the space of spline parameters, autonomous systems can generate motions that are inherently smooth, kinematically feasible, and computationally efficient.
The Representation Gap
The core problem in robotics is often framed as finding a path through a graph or a grid. However, as Jean-Claude Latombe noted in early work on car-like robots, an unconstrained path is useless if the vehicle cannot follow it. Traditional planners often produce "robotic" or jagged paths that must be refined.
Hongliang Dai demonstrates this in EB-RRT*, where raw RRT* outputs are post-processed using cubic Bezier curves. This transformation ensures that the path respects the minimum turning radius of a UAV. Similarly, in the realm of human-like NPC behavior, Ilya Makarov and colleagues use splines to soften mechanical grid-based movements into fluid, believable trajectories.
Splines as a Compressed Search Space
Beyond mere smoothing, splines offer a powerful way to compress the optimization problem. A trajectory spanning several seconds might require hundreds of discrete waypoints to describe accurately, but it can often be represented by just a handful of B-spline control points.
In PE-PSO, Minze Li and Wei Zhao use this dimensionality reduction to slash drone swarm planning times from seconds to milliseconds. Because the optimizer only needs to tweak a few control points to reshape the entire curve, the search space is significantly smaller. Julius Jankowski leverages a similar "via-point" parameterization in VP-STO to enable high-DOF robot arms to react to dynamic obstacles in real-time.
Differentiability and Gradient-Based Planning
The most significant advantage of spline-based representation is differentiability. Because a B-spline is a well-defined mathematical function, we can calculate its derivatives (velocity, acceleration, and jerk) analytically. This allows planners to use gradient-based optimization to "pull" the trajectory away from obstacles while "tightening" it to improve speed or fuel efficiency.
Mohak Bhardwaj and Nathan Ratliff demonstrate this in STORM, a GPU-parallelized MPC framework. By sampling in the space of B-spline coefficients, STORM ensures that every candidate trajectory is smooth by construction, allowing the robot to perform reactive manipulation at 125Hz. Similarly, Shangjie Xue uses Neural Field Dynamics to optimize B-spline parameters via gradient descent, enabling a robot to manipulate granular piles by following complex, curved paths.
Ensuring Second-Order Continuity
For physical hardware, the "smoothness" of a path is often defined by its $C^2$ continuity—meaning its velocity and acceleration are continuous. Discrete jumps in acceleration (infinite jerk) can cause motors to saturate or vibrate.
Zhaoliang Zheng and Thomas Bewley emphasize this in BTO-RRT, where cubic splines are used to ensure that the robot's mechanical stability is never compromised by sudden changes in force. This is equally critical in autonomous underwater vehicles, where Muhammad Ali shows that combining deep reinforcement learning with Non-Uniform Rational B-splines (NURBS) allows underactuated vehicles to follow paths that are physically achievable in chaotic ocean currents.
Integrating Curves into the Search
The latest frontier in this research direction is moving the spline interpolation from a post-processing step into the heart of the search algorithm itself. An Thai Le argues in Global Tensor Motion Planning and Motion Planning Diffusion that integrating splines directly into the sampling or generative process eliminates the need for expensive secondary smoothing.
By representing trajectories as B-spline control points within a generative diffusion model, Le shows that robots can learn to "dream" smooth, multimodal trajectories that are differentiable by construction. This shift suggests that the future of motion planning lies not in finding the best sequence of points, but in optimizing the best continuous function.