Learning Parameterized Maneuvers from Multiple Demonstrations

Hand-specifying trajectories for control tasks can be very time-consuming, and often near impossible as good target trajectories for control should satisfy the system dynamics. This book has concerned with the problem of planning arbitrary trajectories by stitching together small pieces of different parameterized maneuvers. The maneuvers are found by using interpolation-based algorithms and probabilistic model-based algorithms. An algorithm is presented which uses a few waypoints with partial state information and a large corpus of random demonstrations. It then plans a larger trajectory by looking up good demonstrations in the data set based on the waypoints, and then interpolating between the demonstrations picked. This makes is possible to automatically generate target trajectories for control by learning parameterized maneuvers from multiple demonstrations of the maneuvers. As test platform a Drift-R Sedan 4WD 1/10 RC car has been used. A Differential Dynamic Programming controller have been used for successfully controlling the car around the planned trajectory.