Learning-based Motion Stabilizer Leveraging Offline Temporal Optimization

Abstract

During loco-manipulation, instabilities to the robot’s base can be introduced by the manipulator’s motions. Trajectories that are generated on-the-fly may jeopardize the stability and safety of the robot and its surroundings. This work proposes a self-supervised learning-based pipeline to keep a robot stable while executing a given trajectory. Empirical results show that the desired objective can be achieved with the proposed pipeline. Experiments are done in simulation and on hardware on a unique multi-modal, manipulation-capable legged robot, and its scalability is tested on a conventional manipulator.