Convex-Neural RRT* cuts robot path planning time by up to 98%

Sampling-based path planners like RRT* are widely used in robotics for their probabilistic completeness, but they often require many iterations to find high-quality paths. A new paper from Hichem Cheriet and colleagues introduces Convex-Neural RRT*, an algorithm that supercharges the classic approach with learned guidance. The system first uses a neural network to predict informative waypoint regions near likely high-quality paths, then extracts convex candidate regions from those predictions. This lets the planner concentrate exploration on geometrically relevant areas while still maintaining global coverage.

To validate their method, the team ran extensive experiments across three environment types and 18 benchmark maps, comparing Convex-Neural RRT* against Neural RRT*, Neural Informed RRT*, classical RRT*, and LTA*. The results are striking: computation time dropped 30–75% compared to other neural-guided variants, and a massive 88–98% relative to LTA*. At the same time, path length was reduced by an average of 5% versus classical RRT*, with even larger improvements in complex environments. The algorithm also achieved an overall success rate above 99% across varying obstacle densities.

The core innovation lies in combining neural prediction with convex geometry. Instead of wasting samples on the entire space, the neural network identifies promising regions, and the convex extraction ensures the sampling remains mathematically well-behaved. This hybrid approach avoids the common pitfalls of purely neural methods—like overfitting to training environments—by maintaining a principled exploration strategy. The result is a planner that is both fast and robust, capable of handling diverse obstacle configurations without retraining.

For practical robotics, this means faster, higher-quality path planning for time-sensitive applications such as autonomous delivery, search-and-rescue, and industrial automation. The paper provides a strong argument that learning-guided sampling doesn't have to sacrifice reliability for speed. By achieving 98% time savings in some cases while keeping path quality high, Convex-Neural RRT* sets a new benchmark for efficient robot navigation. The code and data are expected to be released alongside the paper on arXiv.