Vehicle Guidance in Dynamic Environments based on Pattern Recognition of Quasilocal Spacetime Embeddings

Arti cial potential elds APFs are well known for industrial robot path plan ning when static obstacles are present Khatib O line supervision and intervention at several stages of the global trajectory design process is possible and often necessay because of spurious local minima The method runs into problems however for the unsupervised case of real time autonomous vehicle control within complex rapidly time varying environments APFs as well as impedance control may yield trajectories that are arbitrary far away from the desired ones or the solutions found do not respect the dynamics constraints of the vehicle In real time applications often there will be no time left to invoke computationally intensive replanning when heuristics fails Control laws for complex time variant geometric con gurations tend to be much more complex then APFs In order to be e cient they must exploit short term predictable aspects of the world From this motivation derives the modular planner design presented in this paper As common in control engi neering we separate the problem into two parsts rstly we form some quasilo cal space time spreadout of the estimated future obstacle trajectories secondly we compute an action policy for vehicle guidance from this by means of some parsimonious high order net parsiHON which has universal approximation capabilities Fahner a