GraphDistNet: A Graph-based Collision-distance Estimator for Gradient-based Trajectory Optimization

TL;DR

GraphDistNet is a novel graph-based neural network that efficiently predicts collision distances and gradients, improving the safety and accuracy of trajectory optimization in complex environments.

Contribution

It introduces GraphDistNet, a new collision-distance estimator using edge feature convolution and attention mechanisms for better generalization and efficiency in trajectory optimization.

Findings

Outperforms baseline methods in simulated environments.

Effective in real-world trajectory optimization tasks.

Generalizes well to unforeseen complex geometries.

Abstract

Trajectory optimization (TO) aims to find a sequence of valid states while minimizing costs. However, its fine validation process is often costly due to computationally expensive collision searches, otherwise coarse searches lower the safety of the system losing a precise solution. To resolve the issues, we introduce a new collision-distance estimator, GraphDistNet, that can precisely encode the structural information between two geometries by leveraging edge feature-based convolutional operations, and also efficiently predict a batch of collision distances and gradients through 25,000 random environments with a maximum of 20 unforeseen objects. Further, we show the adoption of attention mechanism enables our method to be easily generalized in unforeseen complex geometries toward TO. Our evaluation show GraphDistNet outperforms state-of-the-art baseline methods in both simulated and real world tasks.