Proprioceptive and Kinesthetic Feedback for Robotic Manipulation

Abstract

Programming robots to manipulate general objects in unstructured, human-centric environments is a complex problem. Typical assumptions for engineering-based solutions to manipulation are that the object can be recognized, that it is nearly rigid, and that its pose (position and orientation) can be estimated. Even under these stringent assumptions, or perhaps because of them, autonomous task performance has remained primitive over decades of research. The reasons for the slow pace of advancement are generally unknown, however careful engineering of robotic systems to perform one-off tasks i.e., without generalizability) continues.

Autonomous manipulation (i.e., interacting with the environment to accomplish a task) remains one of the biggest challenges roboticists face. Manipulation breakthroughs will happen with better sensing technologies, but we can also better use some of the sensing capabilities already available to robots (e.g., proprioception and kinesthesis) to improve manipulation abilities.

To decipher the importance of those sensations we have examined what humans do within the context of specific tasks (e.g., manipulating bulky or articulated objects), we have extracted human-inspired strategies, and we have demonstrated their use on robotics platforms to achieve the same task with limited sensorial feedback. We have shown the use of proprioception to address the problem of equalizing bulky loads for which caging and closure manipulation techniques are difficult or impossible to apply. Motivated by the utility we discovered in ``alternative'' sensing modalities on this application, we explored the use of proprioceptive and kinesthetic feedback in other areas, like manipulation of articulated objects. We conducted a human study to discover the feasibility of these sensing modalities on a complex manipulation task in the absence of both tactile and visual feedback and demonstrated the ability of a robotic platform to build kinematic models of articulated objects by exclusively using proprioceptive and kinesthetic feedback, similarly to the study subjects.