Inferring the Effects of Wiping Motions based on Haptic Perception

From Control Systems Technology Group

Jump to: navigation, search

Service robots are envisaged as universal assistant in everyday environments. As such, they will have to cope with a wide variety of daily household chores. Within the context of robotic manipulation, but especially for cleaning tasks, it is important to monitor the performance of the executed actions. Typically, humans evaluate the outcome of their actions based on visual perception. However, visual feedback is an unreliable source of information to evaluate some of the most frequent cleaning tasks since dust and streaks of water are hardly perceivable on camera images. It turns out that humans do not solely rely on perceptual feedback but also use abstract process models to maintain knowledge of their actions and effects. The main feedback in this case is haptic information occurring from contact. From this haptic feedback, a human can determine that contact with the to-be-cleaned surface is made and rate its task performance by comparing the desired contact force and the perceived contact force. To this end, the combination of cognitive capabilities and haptic perception enables humans to qualitatively reason about the effects of their motions and solve even complex cleaning tasks despite poor visual perception. Accordingly, it is proposed to utilize the torque sensing capabilities of compliant light weight robots to infer contact situations of compliant wiping motions and measure the task performance based on a qualitative effect model.

This work is based on the representation and planning methods of another source. These include (i) an approach to detect contact during real world wiping motions, (ii) inference methods to estimate the performance of these motions, (iii) a probabilistic contact model to incorporate different contact situations arising from different tool-medium-surface constellations, and (iv) an approach to distinguish “good” contact situations from “bad” contact situations. Furthermore, (v) planning algorithms from the source enable to replan additional wiping motions to enhance the cleaning result w. r. t. bad contact situations introduced by external disturbances.

After a short review of the state-of-the-art, a particle distribution model and outline of the planning algorithms to compute efficient cleaning motions is introduced. This particle distribution model was utilized to estimate the effect of robotic cleaning actions with reference to haptic feedback information. Next, the effect inference strategy to estimate the task performance of wiping motions in a cleaning scenario was introduced and the approach on failure detection and recovery is presented. A log-likelihood based contact model allows to simulate different tool-medium-surface constellations. The method is able to infer failure situations due to bad contact arising from human intervention. It is shown that the robot can plan additional wiping motions based on the inferred information in order to successfully accomplish the commanded tasks despite prior failure situations.

Personal tools