Learning Vision-based Reactive Policies for Obstacle Avoidance

Elie Aljalbout, Ji Chen, Konstantin Ritt, Maximilian Ulmer and Sami Haddadin

Abstract: In this paper, we address the problem of vision-based obstacle avoidance for robotic manipulators. This topic poses challenges for both perception and motion generation. While most work in the field aims at improving one of those aspects, we provide a unified framework for approaching this problem. The main goal of this framework is to connect perception and motion by identifying the relationship between the visual input and the corresponding motion representation. To this end, we propose a method for learning reactive obstacle avoidance policies. We evaluate our method on goal-reaching tasks for single and multiple obstacles scenarios. We show the ability of the proposed method to efficiently learn stable obstacle avoidance strategies at a high success rate, while maintaining closed-loop responsiveness required for critical applications like human-robot interaction.

System Overview

At a certain time step, the system receives a visual input o, joint angle measurements q and joint velocities q˙ . The image o is then encoded using a β-VAE encoder to produce the visual latent code z. Using the robot kinematics φ and jacobian J, we obtain the end-effector position x and velocity x˙ from q and q˙ respectively. Subsequently q and q˙ are fed into a baseline policy to produce a user-defined behavior. Simultaneously, we feed all available information s_r into a reactive policy π_ψ. The latter produces a reactive behavior dependent on the objects in the environment. Both outputs are then composed together based on the RMPflow framework, to produce a desired joint acceleration q¨_d . This acceleration is then fed into the robot controller

Video