Abstract
This paper presents a new approach to learning for motion planning (MP) where critical regions of an environment are learned from a given set of motion plans and used to improve performance on new environments and problem instances. We introduce a new suite of sampling-based motion planners, Learn and Link. Our planners leverage critical regions to overcome the limitations of uniform sampling, while still maintaining guarantees of correctness inherent to sampling-based algorithms. We also show that convolutional neural networks (CNNs) can be used to identify critical regions for motion planning problems. We evaluate Learn and Link against planners from the Open Motion Planning Library (OMPL) using an extensive suite of experiments on challenging motion planning problems. We show that our approach requires far less planning time than existing sampling-based planners.
Introduction
Motion planning—the task of finding a collision-free path for a robot from a start to a goal configuration—is a computationally intensive, NP-hard problem. Modern robotics relies heavily on sampling-based planners like Rapidly-exploring Random Trees (RRT) or Probabilistic Roadmaps (PRM). These algorithms typically use uniform sampling to build a connectivity map of safe paths.
However, uniform sampling often fails in environments with "narrow passages" or bottlenecks. Because these critical regions occupy a tiny fraction of the total configuration space, the probability of randomly sampling a point within them is extremely low, leading to planning failures or excessive computation times. This paper introduces "Learn and Link" (LL), a framework that uses deep learning to identify these bottlenecks from top-down environmental images, allowing planners to focus their efforts where they are needed most.
Methodology
Learning Critical Regions
The core of the approach is a fully convolutional encoder-decoder network (similar to SegNet) that performs pixel-wise classification on 2D images of the environment. The network is trained to predict "critical regions"—areas where successful trajectories frequently pass but are unlikely to be sampled uniformly.
Automated Label Generation
To train the network without manual annotation, the authors developed an automated pipeline. They ran an off-the-shelf RRT-Connect planner multiple times on a set of training environments to generate overlapping motion traces. By applying Itti's visual saliency model to these traces, they could automatically identify high-saliency bottleneck regions to serve as ground-truth labels for the CNN.
The Learn and Link (LL) Planners
The framework introduces two specific planners: LLP (for single-query tasks) and LL-RM (for multi-query roadmaps). These planners seed their roadmaps with points drawn from the CNN-predicted critical regions.
A key insight of this work is that simply seeding points is insufficient; local planners often lack the robustness to connect these isolated bottleneck points to the wider open space. To address this, LL treats each initial configuration as the root of its own subgraph and uses an RRT-Connect-style tree expansion to actively grow and link these subgraphs together. Importantly, because the system maintains a base level of uniform sampling, it preserves the "probabilistic completeness" of traditional planners—the guarantee that a solution will be found if one exists, even if the neural network's predictions are suboptimal.
Experimental Results
The LL planners were evaluated on SE(2) mobile base tasks and high-dimensional 10-DOF tasks (a Barrett WAM arm on a mobile base), comparing performance against standard algorithms in the Open Motion Planning Library (OMPL).
- Efficiency: For successful plans, LL planners required 88% to 99% less planning time on average than the best-performing standard OMPL planners (such as RRT-Connect or PRM).
- Reliability: In complex environments with tight corridors where standard planners frequently had a 0% success rate within a 60-second limit, LL planners achieved 99-100% success rates.
- Practicality: CNN inference takes less than 5 seconds from a top-view image, significantly faster than competing Variational Autoencoder (VAE) approaches that may require up to 50 seconds for data structuring alone.
Conclusion and Future Work
Learn and Link demonstrates that combining deep learning with traditional sampling-based architectures can yield massive performance gains while maintaining mathematical guarantees. While the current implementation relies on 2D top-down representations, the results suggest a powerful path forward for motion planning. Future research will focus on extending critical region prediction directly into higher-dimensional configuration spaces or 3D voxel grids to better support complex manipulator arms and non-planar environments.
Summary updated