Deep Coverage: Learning and Control for Compliant Robots

Robots have thrived in highly structured, accurately known, and safely enclosed industrial settings, as in the left figure below (from Kuka robotics). Typically, these platforms consist of rigid elements connected by a few degrees of freedom (DoF) that can be controlled directly. This allows strict enforcement of motion for high speed, accuracy, and consistency of repeated tasks. In contrast, modern robotics aspires to more general and ambitious operational goals involving highly unstructured environments, arbitrary terrains, unknown objects, or the company of humans. This means that traditional objectives are superseded by fundamental needs for adaptability and safety.

These altered priorities motivate robots with versatile physical features, such as deformation and heightened articulation, which can be achieved by using soft components and a high number of compliant DoF, as in the center and right figure above (soft hand by Deimel and Brock, tensegrity robot structure from NASA Ames). These features allow for adaptive contact geometry and storage or dissipation of energy for purposes such as efficient mobility on rough terrain and manipulation of arbitrary objects. Examples include soft and adaptive hands, bio-inspired modular robots, underactuated legged systems, as well as tensegrity rovers, dynamic truss structures that incorporate length-actuated members.

While such platforms provide increased safety and adaptability, they pose many complications to the prediction and regulation of motion. Due to higher system dimensionality and imperfect sensing of environmental variation, natural dynamical influences such as momentum and contact forces become both more significant and more complex to describe. Nevertheless, a pivotal requirement is to achieve motion coverage with such robots. This expresses the need to effectively command motion for a wide range of operational bounds. This work aims to outline a pathway toward this goal in light of ongoing progress in motion synthesis and data-driven design.

Check the following Blue Sky paper regarding the proposed idea of "deep coverage":

Surovik, D., and KE Bekris. 2017. “Deep Coverage: Motion Synthesis In The Data-Driven Era”. In International Symposium on Robotics Research (ISRR), Puerto Varas, Chile.