SPADE: A spatial information assisted collision distance estimator for robotic arm

Abstract

The movement of a robotic arm in the working environment requires efficient and adequate motion planning. The procedure of collision detection based on the object geometry is crucial to plan the motion trajectories, and usually requires intensive resource and considerable time. Many learning-based collision detection schemes have been developed to improve the efficiency of collision detection. However, current learning-based collision detection methods are either not accurate enough or prone to low efficiency. We propose a simple, yet highly accurate collision distance estimator, a spatial information assisted distance estimator, i.e., SPADE, in which spatial information of both robotic arms and obstacles are encoded by multiple encoders. With evaluation in both static and dynamic environments, our model shows higher prediction accuracy than multiple baselines, and higher accuracy can be corroborated by experiment with our model under the premise of equal inference efficiency. In addition, our model shows better robustness than baseline in real-world path planning.