Learning and Prediction Method for Collision Distance Prediction Models, and Apparatus Therefor

Real-time Motion Planning for Robots

This technology extracts individual components of robots and obstacles, then predicts collision distances in parallel through pairwise batch operations. It trains a collision distance prediction model based on geometric feature vectors and relative transformation matrices. The minimum value among the predicted pairwise distances is calculated as the global collision distance, which can then be utilized for real-time motion planning.

Previously, high computational complexity led to performance degradation when calculating minimum distances, a crucial step for conventional motion planning algorithms in high-degree-of-freedom robot systems. Furthermore, data-driven learning methods suffered from low flexibility to environmental changes and frequent retraining requirements, limiting their versatility.

This technology proposes a model that learns by extracting relative transformation values and point cloud-based shape feature vectors between robot components and obstacles. By processing these inputs in batches and performing parallel computations, it enhances operational efficiency and provides flexibility to adapt to environmental changes without needing to retrain for specific shape elements.

‍

‍

Key Features:

• The training device for the collision distance prediction model extracts all movable components of robots and obstacles.
• Inputs component pairs, which are formed by grouping each extracted obstacle component into pairs.
• Predicts the pairwise collision distance for each component pair, representing the minimum distance between the components in that pair.
• Trains the collision distance prediction model using a loss function that compares the predicted global collision distance with the actual collision distance.

‍

This technology was developed with support from the Institute of Information & Communications Technology Planning & Evaluation (IITP) through its goal-oriented AI generation and inference research project.

‍