Collision detection is essential to virtually all robotics applications. However, traditional geometric collision detection methods generally require pre-existing workspace geometry representations; thus, they are unable to infer the collision detection function from sampled data when geometric information is unavailable. Learning-based approaches can overcome this limitation. Following this line of research, we present DeepCollide, an implicit neural representation method for approximating the collision detection function from sampled collision data. As shown by our theoretical analysis and empirical evidence, DeepCollide presents clear benefits over the state-of-the-art, as it relates to time cost scalability with respect to training data and DoF, as well as the ability to accurately express complex workspace geometries. We publicly release our code.

翻译：碰撞检测对几乎所有机器人应用都至关重要。然而，传统的几何碰撞检测方法通常需要预先存在的工作空间几何表示；因此，当几何信息不可用时，它们无法从采样数据推断碰撞检测函数。基于学习的方法可以克服这一局限性。遵循这一研究方向，我们提出了DeepCollide，一种用于从采样碰撞数据近似碰撞检测函数的隐式神经表示方法。正如我们的理论分析和实证证据所示，DeepCollide相较于现有技术展现出明显优势，具体体现在训练数据与自由度的可扩展性时间成本，以及精确表达复杂工作空间几何形状的能力。我们公开发布了代码。