Abstract
This paper utilizes finite Fourier series to represent a time-continuous motion and proposes a novel planning method that adjusts the motion harmonics of each manipulator joint. Primarily, we sum the potential energy for collision detection and the kinetic energy up to calculate the Hamiltonian of the manipulator motion harmonics. Though the adaptive interior-point method is designed to modify the harmonics in its finite frequency domain, we still encounter the local minima due to the non-convexity of the collision field. In this way, we learn the collision field through a support vector machine with a Gaussian kernel, which is highly convex. The learning-based collision field is applied for Hamiltonian, and the experiment results show our method's high reliability and efficiency.
Introduction
Motion planning for robotic manipulators requires a delicate balance between trajectory smoothness and collision safety. Traditional planners often struggle with two primary issues: local minima in non-convex obstacle fields and discretization errors. Numerical optimization methods like CHOMP or TrajOpt are efficient but frequently get trapped in local minima where the gradient fails to guide the robot out of a collision. Conversely, sampling-based methods like RRT* are more robust but computationally intensive and often produce jerky paths.
Furthermore, most planners rely on discrete waypoints. This can lead to "stuck in time" scenarios where the robot appears safe at specific points but undergoes a collision during the transition between them. This paper introduces FFS-OMP, a framework that addresses these challenges by using continuous trajectory representation and a learned convex collision field.
Methodology
The FFS-OMP framework is built upon three core innovations designed to ensure smoothness and global-like convergence in local optimization.
Finite Fourier Series (FFS) Trajectories
Instead of discrete waypoints, the robot's trajectory is represented as a sum of sine and cosine harmonics. This mathematical formulation ensures the motion is inherently time-continuous and smooth. To address potential discontinuities at the start and end points of a standard Fourier series, the authors employ a Cosine Series approach. By symmetrizing the trajectory, the Cosine series eliminates the "jumps" often seen at the boundaries of periodic functions, providing a more stable basis for optimization.
Learning-based Collision Field (SVM)
To overcome the "stuck in space" problem caused by non-convex obstacles, the authors utilize a Support Vector Machine (SVM) with a Gaussian kernel. By training on 3D Euclidean space data (labeled as safe or collided), the SVM transforms a complex, jagged environment into a smooth, highly convex potential field. This convexity allows the optimizer to effectively "see" a clear path out of collisions that would typically trap standard gradient-descent algorithms.
Adaptive Interior Point (AIP) Optimization
The planning problem is framed as minimizing a Hamiltonian—the sum of kinetic energy (for smoothness) and potential energy (for safety). The authors implement an "Adam-like" adaptive gradient approach within an interior-point method. This uses exponential moving averages to dampen oscillations, preventing the robot from vibrating between the competing requirements of smoothness and obstacle avoidance.
Experimental Results
The method was evaluated on 7-DoF LBR-iiwa and 6-DoF AUBO-i5 robots across various difficulty classes. In the most challenging scenarios (Class C), FFS-OMP maintained a success rate of 95–96.75%, whereas traditional planners like TrajOpt and GPMP2 often fell below 15%.
In terms of efficiency, FFS-OMP significantly outperformed sampling-based alternatives. On the LBR-iiwa, it averaged approximately 2.3 seconds per plan, compared to 13.6 seconds for RRT-Connect. Additionally, the SVM-based collision field provided a 45% acceleration in optimization convergence compared to standard numerical baselines.
Conclusion
FFS-OMP demonstrates that combining continuous harmonic representations with learned convex collision fields significantly improves the reliability and speed of motion planning. While currently focused on static environments, the method's ability to escape local minima and ensure time-continuous safety marks a substantial advancement over waypoint-based numerical optimizers. Future work may explore real-time updates for dynamic environments and further hyperparameter automation.
Summary updated