Neural-Based Self-collision Checking for a Quadruped Robot

Motion planning of legged robots requires constraints checking. The primary constraint is related to the robot’s kinematic model and self-collision checking. Checking this constraint allows the full utilization of the robot workspace during motion planning. The most popular self-collision checking methods utilize a 3D mesh model of the robot and iterative methods to find colliding parts of the robot. This approach is accurate but slow, so in this paper, we study the application of Multilayer Perceptron to build a 3D self-collision model of the legged robot. We employ the neural network to classify the state of the robot into two collision and collision-free binary values. A similar approach has already been applied for manipulating robots, but legged systems are defined in higher-dimensional space, so the problem is significantly more challenging. In this paper, we study the influence of input to the model on the neural network performance. Then, we show that the application of Fourier features enhances the input vector and improves the classification results. We demonstrate the results on the model of the quadruped walking robot ANYmal C.