An overview of three outstanding systems for motion planning based on graphical neural networks

Graph Neural Networks (GNN)-based motion planning has emerged as a promising approach for robotic systems due to its efficiency in pathfinding and navigation tasks. This approach uses GNNs to learn the underlying graph structure of an environment to make quick and informed decisions about which paths to take. Let’s look at the detailed specifics of the three main systems:

1. GraphMP: A graph neural network based motion planner

GraphMP is a neural motion planner designed for tasks of varying dimensionality, from 2D mazes to high-dimensional robotic arms. The main strength of GraphMP lies in its tailored architecture and training mechanism, which facilitates efficient graph pattern extraction and graph search processing.

Architecture and training:

  • Collision checker: This GNN-based module detects obstacles by analyzing the graph structures of the environment and can therefore efficiently predict potential collisions.
  • Heuristic estimator: This component helps refine the graph search for optimal paths by estimating the path cost.

The system uses an end-to-end training approach that allows it to recognize chart patterns while performing chart searches.


  • GraphMP consistently outperforms classic planners (like A*) and state-of-the-art learning-based planners on tasks like navigating a 14D robotic arm.
  • Its unique model architecture and training approach significantly improved path quality and planning speed.


  • From the 2D maze to the 14D dual KUKA robot arm: GraphMP has significantly improved path quality and scheduling speed compared to existing planners.
  • Success rate: Almost 100% success rate in different environments, proving its adaptability.


  • Path quality: Up to 25% better than the competition.
  • Planning speed: Up to 40% faster than traditional planners.
  • Success rate: Almost perfect for most tasks.

2. End-to-end neural movement planner

This planner emphasizes safety and compliance in urban environments. Integrating LIDAR data and HD maps generates detailed 3D representations and predictions for self-driving cars.


  • Uses a convolutional network backbone to calculate cost volumes that evaluate potential paths.
  • The model processes raw LIDAR data and maps and generates intermediate representations such as 3D recognition and trajectory predictions.


  • Cost volumes guide trajectory capture, helping the car navigate safely by minimizing potential collisions.
  • Trained with a multitasking objective focused on planning, detection and path optimization.


  • Demonstrated effectiveness in complex urban environments and demonstrated ability to adapt to real-world driving scenarios.
  • Outperformed leading neural architectures in 3D recognition and motion prediction accuracy.


  • Detection accuracy: Outperformed leading neural architectures.
  • Flight path safety: Minimizes collision risks by complying with traffic rules.
  • Planning speed: Real-time trajectory planning enables safe navigation.

3. Motion Planning Networks (MPNet)

MPNet integrates deep learning into motion planning to efficiently navigate high-dimensional spaces. Its encoder network creates a latent spatial representation of the obstacles and its planning network predicts paths based on the robot’s configuration.


  • It uses an encoder network to convert point cloud data into a latent space.
  • The planning network uses this information to predict paths based on the robot’s configuration.


  • The point cloud encoder and planning network map obstacles and predict collision-free paths.
  • Combines neural planning with traditional motion planning (RRT*) to robustly handle complex planning tasks.


  • MPNet transfers well to unfamiliar environments and has robust adaptability.
  • Keeps execution times under one second in various scenarios.


  • Execution time: Less than a second in most scenarios.
  • Success rate: 85% success rate in challenging high-dimensional environments.
  • Path quality: Paths are optimized based on the latent space encoded by the network.

Comparison table


Graphical neural network-based motion planning offers significant advances in robot navigation. The different approaches of GraphMP, the end-to-end planner and MPNet show that this technology can be adapted to a variety of environments and offers speed, efficiency and security in planning optimal paths for autonomous systems.


Sana Hassan, Consulting Intern at Marktechpost and dual degree student at IIT Madras, is passionate about using technology and AI to address real-world challenges. With a keen interest in solving practical problems, he brings a new perspective to the interface between AI and real-world solutions.

Source link