OctoPath: An OcTree-Based Self-Supervised Learning Approach to Local Trajectory Planning for Mobile Robots.

Area of Science:

• Robotics
• Artificial Intelligence
• Computer Vision

Background:

• Autonomous mobile robots face challenges in complex environments, requiring effective obstacle recognition, path planning, and motion execution.
• Current trajectory estimation methods often rely on regression, leading to an infinite output state space and potential inaccuracies.

Purpose of the Study:

• To introduce OctoPath, a self-supervised encoder-decoder deep neural network for predicting local optimal trajectories for autonomous mobile robots.
• To reformulate trajectory prediction as a classification problem using a 3D octree environment model for improved accuracy and efficiency.

Main Methods:

• OctoPath utilizes a 3D octree discretization for environment modeling, enabling trajectory prediction as a classification task.
• The network is trained in a self-supervised manner, minimizing the error between predicted and manually driven trajectories.
• Sensor fusion combines data from a 40-channel mechanical LiDAR, an inertial measurement unit, and wheel odometry for state estimation.

Main Results:

• OctoPath demonstrated effective local optimal trajectory prediction in both simulated and real-world driving scenarios.
• The system was benchmarked against hybrid A-Star, regression-based supervised learning, and CNN-based methods, showing competitive or superior performance.
• Experiments were conducted indoors and outdoors, validating the robustness of OctoPath across diverse environments.

Conclusions:

• OctoPath offers a robust and accurate solution for local trajectory prediction in autonomous mobile robot navigation.
• The self-supervised, classification-based approach effectively overcomes limitations of traditional regression-based methods.
• The system shows promise for enhancing the safety and efficiency of robots operating in complex, dynamic environments.