Neural Implicit Swept Volume Models for Fast Collision Detection

TL;DR

This work presents a novel neural implicit swept volume model to continuously represent arbitrary motions parameterized by their start and goal configurations that allows to quickly compute signed distances for any point in the task space to the robot motion.

Abstract

Collision detection is one of the most time-consuming operations during motion planning. Thus, there is an increasing interest in exploring machine learning techniques to speed up collision detection and sampling-based motion planning. A recent line of research focuses on utilizing neural signed distance functions of either the robot geometry or the swept volume of the robot motion. Building on this, we present a novel neural implicit swept volume model to continuously represent arbitrary motions parameterized by their start and goal configurations. This allows to quickly compute signed distances for any point in the task space to the robot motion. Further, we present an algorithm combining the speed of the deep learning-based signed distance computations with the strong accuracy guarantees of geometric collision checkers. We validate our approach in simulated and real-world robotic experiments, and demonstrate that it is able to speed up a commercial bin picking application.

Figures (6)

• Figure 1: The basic idea of our approach is to predict the signed distance (black line) of a point in the task space (black dot) to its nearest point (white dot) on the surface of the swept volume (blue mesh) of a robot motion. The swept volume is implicitly represented by a neural network, which takes the start and goal configuration of the motion as well as the query point as inputs and outputs the signed distance of the point to the implied swept volume. The robot motion can be collision checked against a complete environment, by representing the environment as a point cloud. Each of these points becomes one input point for the network.
• Figure 2: Network architecture: each blue rectangle represents one of the $n_{\text{b}}$ blocks of the network. The input (green) of the network is fed into each block individually. The gray layers are linearly mapping the input and output dimension of the network to the constant block dimension $n_{\text{dim}}$.
• Figure 3: Exemplary comparison of different model sizes for the linear joint space motion already shown in Fig. \ref{['fig:intro']}. There is a noticeable degradation when going from (c) 11x512 to (d) 5x512, while the difference between (c) 11x512 and (b) 11x1024 is less pronounced. This qualitative finding is also reflected in the metrics shown in Table \ref{['tab:mae_lin_spline']}.
• Figure 4: Comparison between the ground truth swept volume meshes (gray) and meshes reconstructed from the neural networks (red) via the marching cubes algorithm (resolution: 1 cm). The upper row shows linear joint space motions, while the lower row shows spline motions. For both rows, 11x1024 models were used to predict the signed distances. The motions are the first four motions of their respective test sets, and thus have not been used in training.
• Figure 5: An example of a box world scene, with randomly sampled boxes and points on the surface of the boxes.