Our paper titled “Guaranteed obstacle avoidance of convex objects in 2D without object tracking, environment mapping, or stopping” has been accepted to the IEEE Control Systems Letters journal.

The paper shows that a simple end-to-end controller enables collision avoidance behaviors with guarantees when using 2D LiDAR for feedback. Most approaches with guarantees rely on tracking objects, modeling the environment (mapping), or stopping the robot as the distance to the nearest object decreases. All these have shortcomings which we avoid. The resulting controller is computationally very cheap, and comes with some guarantees.

The core idea is to develop a barrier function for relative motion kinematics between robot and object expressed in an object-boundary-centric frame. The barrier function has a relative-degree of one, and its validity condition provides a constraint on the velocities of the robot. We then show that the proposed end-to-end controller, despite having no state estimation, ends up meeting this condition for any convex object in certain situations, guaranteeing safety.

The weakness of the method is that the situations where the guarantee works leaves out head-on-collisions, like when the robot is moving towards a wall orthogonal to it. Any slight deviation from head-on situation will result in guaranteed collision avoidance.