Goal-Predictive Robotic Teleoperation using Predictive Filtering and Goal Change Modeling

Robotic teleoperation involves input sensors which can be used to operate mobile robots. We have used Microsoft Kinect depth camera as our input sensors in our goal-predictive robotic teleoperation systems. Since Kinect depth camera is inexpensive but has a lot of errors in human skeleton tracking, especially position noise as it is difficult to track human positions in real-time. We have tried to de-noise the Kinect outputs using Inverse Optimal Control for Linear Quadratic Regulator (LQR) using current time step data for intended goal prediction. But there is some amount of log loss in the Inverse LQR goal prediction. Using predictive filters like linear regression has improved the goal prediction and decrease the log loss for true goal. Sometimes human also changes their intended goal in-between the trajectory or path to the previous intended goal, which makes it more difficult for the robot to cope up with humans. The trajectory followed by the demonstrator and the goal intended can be modeled as Hidden Markov Model (HMM), where goal probability change can be observed at each time step in the trajectory. Goal change model is used to find the probability of intended goal change at any time step in the trajectory, calculated effectively through the forward-backward algorithm.