Learning Vision-based Reactive Policies for Obstacle Avoidance

During task execution, robots should be capable of operating in their workspaces without colliding with obstacles. In well-structured environments, this constraint can be easily ensured by carefully designing collision-free motion trajectories based on the understanding of the robot surroundings. In contrast, unstructured environments present the challenge of autonomously reacting to previously unknown settings. To tackle this challenge, extra efforts are needed in order to design proper perception systems, capable of understanding the environment, as well as reactive strategies to avoid the obstacles. In this work, we are concerned with obstacle avoidance for robot manipulators. In addition to the previously mentioned challenges, such systems impose additional constraints such as joint limits, singularities and self-collision. All of these aspects add to the complexity of the problem, and require proper care in the formulation of both classical and learning-based methods. In this context, proprioceptive robot sensors enable collision detection and early reactions which can prevent substantial damages to the robot and its environment [1].