Safety, stability and learning of haptic teleoperation through control barrier functions
OA Version
Citation
Abstract
Robotic teleoperation enables human operators to engage in activities within dangerous or challenging environments where autonomous control is impractical. This capability is particularly crucial in tasks that demand human intuition and adaptability, such as search-and-rescue operations in intricate environments. However, robotic teleoperation presents challenges for safety and precise navigation, primarily stemming from the typically restricted field of view of the sensors.
Haptic teleoperation emerges as a promising approach to remedy the aforementioned challenges by providing an extra feedback channel through force-based haptics. In this dissertation, we propose a set of innovative approaches to improve the performance of a haptic teleoperation system by leveraging Control Barrier Functions (CBFs), which is a state-of-the-art approach in the area of safety-critical control.
In the first part of the dissertation, we introduce a novel approach that generates force feedback from the difference between a human-issued control input and a safe control input calculated by CBFs. This force feedback helps guide the human operator towards the input command that is closest to their current command and deemed to be safe. This contrasts with previous methods that generally provide force feedback in the opposite direction of an obstacle.
We demonstrate the CBF-based force feedback in two paradigms: Haptic Shared Control (HSC), in which the human user has ultimate control over the robot's actions and the force feedback only provides safety guidance; and Haptic Shared Autonomy (HSA), which guarantees the safety of the robot by overriding the user's input if necessary, and uses the haptic feedback to communicate the discrepancy between the user and the autonomous controller.
Beyond safety, stability stands as another pivotal aspect in haptic teleoperation of robots, particularly for human-in-the-loop systems. In the second part of the dissertation, we present our work for avoiding unexpected and uncontrolled oscillations in a haptic navigation. Specifically, we introduce a differential constraint on the rendered force that makes the system finite-gain L2 stable. Our constraint is related to but less restrictive than the typical passivity constraint used in previous literature. We investigate the small-L2-gain method in both HSC and HSA paradigms, and derive closed-form solutions to the resulting optimization problems to ensure that our method can be implemented in real-time.
Finally, given that the human user engages directly with the force feedback, ensuring the user's satisfaction with the feedback becomes another vital aspect for consideration. In the last section of the dissertation, we introduce a Learning from Haptics (LfH) approach that automatically tunes the CBF constraints based on the user preferences of the force feedback. This learning approach aims to allow the human user to customize the CBF-based haptic feedback in an intuitive and practical manner. Our method takes sparse examples of desired force feedback from the user, and produces new CBF parameters that generate forces that are closer to the desired ones.