Learning Parameterized Maneuvers from Multiple Demonstrations
Author
Sørensen, Martin Møller
Term
10. term
Publication year
2009
Pages
96
Abstract
At specificere målbaner manuelt til styringsopgaver er tidskrævende og svært, fordi banerne skal overholde systemets dynamik. Dette projekt undersøger, hvordan man kan planlægge vilkårlige baner ved at sy små, parametrebestemte manøvrer sammen. Manøverne findes med interpolationsbaserede og probabilistiske, modelbaserede algoritmer. Vi præsenterer en algoritme, der bruger få vejpunkter med delvis tilstandsoplysning (fx position, retning eller hastighed) og et stort korpus af tilfældige demonstrationer. Algoritmen slår gode demonstrationer op baseret på vejpunkterne og interpolerer derefter mellem dem for at danne en længere, glat bane. På den måde kan man automatisk generere målbaner til kontrol ved at lære parametrebestemte manøvrer fra flere demonstrationer. Som testplatform anvendes en Drift-R Sedan 4WD 1/10 RC-bil. En Differential Dynamic Programming-controller (en optimeringsbaseret styringsmetode) bruges til succesfuldt at styre bilen langs den planlagte bane.
Manually specifying target trajectories for control is time-consuming and difficult because trajectories must respect the system’s dynamics. This project explores planning arbitrary trajectories by stitching together small, parameterized maneuvers. The maneuvers are identified using interpolation-based and probabilistic, model-based algorithms. We present an algorithm that takes a few waypoints with partial state information (for example, position, heading, or speed) and a large corpus of random demonstrations. It looks up demonstrations that fit the waypoints and interpolates between them to build a longer, smooth path. This enables automatic generation of target trajectories for control by learning parameterized maneuvers from multiple demonstrations. As a test platform, we use a Drift-R Sedan 4WD 1/10 RC car. A Differential Dynamic Programming controller—an optimization-based control method—successfully controls the car along the planned trajectory.
[This abstract was generated with the help of AI]
Documents