Tactile Guidance for Policy Adaptation

Abstract

Demonstration learning is a powerful and practical technique to develop robot behaviors. Even so, development remains a challenge and possible demonstration limitations, for example correspondence issues between the robot and demonstrator, can degrade policy performance. This work presents an approach for policy improvement through a tactile interface located on the body of the robot. We introduce the Tactile Policy Correction (TPC) algorithm, that employs tactile feedback for the refinement of a demonstrated policy, as well as its reuse for the development of other policies. The TPC algorithm is validated on humanoid robot performing grasp positioning tasks. The performance of the demonstrated policy is found to improve with tactile corrections. Tactile guidance also is shown to enable the development of policies able to successfully execute novel, undemonstrated, tasks. We further show that different modalities, namely teleoperation and tactile control, provide information about allowable variability in the target behavior in different areas of the state space.

Iterative learning of grasp adaptation through human corrections

“…For example, a human teacher might supervise the learning process, by modifying targets learned from demonstration [9] or resolving ambiguities in goal representations [10]. Datasets are iteratively built by providing new demonstrations in areas of low policy prediction confidence [40], [41], by providing explicit corrections on policy predictions to generate new data [40], [12] and by physically touching a robot during execution to provide kinesthetic corrections [11], [42], [13].…”

“…Furthermore, our executions do not depend on time (unlike [11], [42], [13]), as our goal is not to execute a trajectory but rather to respond online to changes in contact with an object.…”

Learning from Demonstration and Correction via Multiple Modalities for a Humanoid Robot

“…The Tactile Policy Correction (TPC) algorithm offers an approach for the adaptation of a demonstrated policy, using tactile feedback from a human teacher [3]. Corrections are provided in order to accomplish two goals (Fig.…”

“…For our initial empirical validations of the TPC algorithm [3], the tactile correction interface consisted of Ergonomic Touchpads encircling the wrist of a manipulator arm, with validation on grasp positioning tasks. Comparisons to policies derived from solely teleoperation demonstration confirmed policy reuse to be an effective mechanism for transferring domain knowledge, and policy refinement to be more successful at improving performance.…”

Continuing Robot Skill Learning after Demonstration with Human Feedback

“…Tactile feedback is used to assist in both policy refinement and the reuse of a demonstrated policy when developing a different policy; effectively using the demonstrated policy as prior knowledge for a new behavior. Empirical validation has included grasp positioning on the iCub humanoid [4], as well as grasp adaptation in response to changes in fingertip contact [13].…”