Intention is what you need to estimate: Attention-driven prediction of goal pose in a human-centric telemanipulation of a robotic hand

This work entails remote telemanipulation of certain objects using Dexmo Haptic glove (DHG) and Allegro Robotic Hand (ARH). We introduce an estimation mechanism to quantify the expected goal pose of fingers of the human user, wearing the DHG, as its intent, defined in terms of the expected rotation angle of the object (about the viewing plane) that is held between the end-effectors of ARH. A significant amount of delay is observed to generate this intent due to communication and control latencies when the robot is remotely controlled. Hence, an attention based mechanism is leveraged to model the trajectory of estimated intent and predict its estimate for a lookahead of 'm' time units from the current n^{th} estimated sample to compensate for the delays. We evaluate the performances of the estimation mechanism, and the attention mechanism on the stated robotic setup in a real-work networking scenario against some benchmark methodologies. The effect of varying lookahead is analysed against the accuracy of estimation/prediction of the intent. The testing MSE achieved in prediction of the human intent (utilizing attention model) is reported to be 0.00047 for m=1, which characterizes as ~38-42 times lesser in comparison to our previous work (utilizing LSTM).