Hybrid hierarchical learning for solving complex sequential tasks using the robotic manipulation network ROMAN

Solving long sequential tasks remains a non-trivial challenge in the field of embodied artificial intelligence. Enabling a robotic system to perform diverse sequential tasks with a broad range of manipulation skills is a notable open problem and continues to be an active area of research. In this work, we present a hybrid hierarchical learning framework, the robotic manipulation network ROMAN, to address the challenge of solving multiple complex tasks over long time horizons in robotic manipulation. By integrating behavioural cloning, imitation learning and reinforcement learning, ROMAN achieves task versatility and robust failure recovery. It consists of a central manipulation network that coordinates an ensemble of various neural networks, each specializing in different recombinable subtasks to generate their correct in-sequence actions, to solve complex long-horizon manipulation tasks. Our experiments show that, by orchestrating and activating these specialized manipulation experts, ROMAN generates correct sequential activations accomplishing long sequences of sophisticated manipulation tasks and achieving adaptive behaviours beyond demonstrations, while exhibiting robustness to various sensory noises. These results highlight the significance and versatility of ROMAN’s dynamic adaptability featuring autonomous failure recovery capabilities, and underline its potential for various autonomous manipulation tasks that require adaptive motor skills.