Skill Transformer: A Monolithic Policy for Mobile Manipulation

Prior works [6, 33, 37] show that long-horizon robotic tasks by combining conditional sequence it is possible to end-to-end learn a single policy that can modeling and skill modularity. Conditioned on egocentric perform such diverse tasks and behaviors. Such a monolithic and proprioceptive observations of a robot, Skill policy directly maps observations to low-level actions, Transformer is trained end-to-end to predict both a highlevel reasoning both about which skill to execute and how to skill (e.g., navigation, picking, placing), and a wholebody execute it. However, scaling end-to-end learning to longhorizon, low-level action (e.g., base and arm motion), using multi-phase tasks with thousands of low-level steps a transformer architecture and demonstration trajectories and egocentric visual observations remains a challenging that solve the full task. It retains the composability and research question.