GraphDistNet: a graph-based collision-distance estimator for gradient-based trajectory optimization

Abstract

Trajectory optimization (TO) aims to find a sequence of valid robot states while minimizing trajectory 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. GraphDistNet can precisely encode the structural information between two geometries, a robot and obstacles, by leveraging edge feature-based convolutional operations, and also efficiently predict of collision distances and gradients by the batch computation. We validate the efficiency and preciseness of our model by estimating $25,000$ random environments with a maximum of $20$ unforeseen objects in the experiment. Further, we show the adoption of message passing neural networks and attention mechanism enables our method to be easily generalized in environments containing unforeseen complex geometries. Our evaluation show GraphDistNet outperforms state-of-the-art baseline methods in both simulated and real world tasks. Part of this work was presented at the 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) and published in 2022 IEEE Robotics and Automation Letters (RA-L).