Teleoperated online Learning from Demonstration in a partly unknown environment

Abstract

The general approach to generate collision free motion in a constraint environment is to use path planners, which demand a known environment and potentially fail otherwise. Learning from Demonstration (LfD) can be used instead to teach the robot unknown parts of the environment, such as a goal deviation or an unforeseen obstacle. The general approach is to train a model offline and expect it to perform well afterwards. Problems arise however when the model is trained insufficiently or unknown variations have occurred in the environment, which demand for refinement of the model. In online learning the operator is allowed to refine a predicted trajectory during execution time, where the state-of-the-art methods focus on kinaesthetic teaching. The contribution of this research is the development of a teleoperated online learningmmethod, where the operator can make refinements by moving a haptic stylus (Phantom Omni) in the desired direction. By doing this, a force is felt proportional to the magnitude of refinement. After creating such a refined trajectory, it is used to update a condition dependent probabilistic trajectory model. The proof of concept was shown on a 2D example and on a simulated robot that shows that we can adapt an initial model when unknown variations occur and that the method is able to deal with different object positions and initial end effector poses. To show if other people can use the method, a human factors experiment is performed, comparing the developed method against three other methods on how much time it takes to successfully adapt a model (refinement time) and on the perceived workload. Two different parameters are varied, which are the teaching device (stylus or keyboard) and the learning mechanism (online or offline). The expectation was that both online with stylus has the lowest refinement time and workload, but the results show that only online has a significant improvement over offline methods (p = 7.94 × 10^−12 and p = 0.000512 respectively). This is explained by the fact that only small corrections have to be made and in a maximum of three degrees of freedom (DoFs). No significant difference was found between keyboard and stylus (p = 0.755 and p = 0.302 respectively). An explanation for this is that this is task, person and implementation dependent. The recommendations are to evaluate the proof of concept on the real robot and to extend the method with orientation refinement, such that more complex tasks can also be dealtwith. We hypothesize that with these tasks the combination of online with stylus does perform the best.