Learning from experience for rapid generation of local car maneuvers

Speeding up deep neural network-based planning of local car maneuvers via efficient B-spline path construction

“…Here machine learning comes to the rescue, as modern methods, like deep neural networks (DNN), make it possible to learn even complicated decision-making policies in constrained state spaces [4]. We have explored this idea in our recent paper [5], where we presented a neural network architecture and a training procedure that allow a local motion planner to learn from its own experience (Fig. 1).…”

“…Although our previously introduced DNN [5] keeps the path computation time below 50 ms, some emergency maneuvers require even faster replanning, at the sensor frame rate of at least 30 fps, still yielding smooth paths satisfying all the constraints. Therefore, we contribute in this paper a novel path parametrization and procedure of its construction, which enables our method to compute yet better paths in an even shorter time in comparison to [5] 1 .…”

“…Although our previously introduced DNN [5] keeps the path computation time below 50 ms, some emergency maneuvers require even faster replanning, at the sensor frame rate of at least 30 fps, still yielding smooth paths satisfying all the constraints. Therefore, we contribute in this paper a novel path parametrization and procedure of its construction, which enables our method to compute yet better paths in an even shorter time in comparison to [5] 1 . The new parametrization, even though it follows the idea of the approximation of an implicitly defined oracle planning function, breaks up with the Markov Decision Process formalism used in [5], instead plans the whole maneuver at once.…”

“…Therefore, we contribute in this paper a novel path parametrization and procedure of its construction, which enables our method to compute yet better paths in an even shorter time in comparison to [5] 1 . The new parametrization, even though it follows the idea of the approximation of an implicitly defined oracle planning function, breaks up with the Markov Decision Process formalism used in [5], instead plans the whole maneuver at once. The path is no longer represented with a sequence of polynomials, but with a single 7th degree B-spline curve.…”

“…We evaluate thoroughly the new planning method using the latest motion planning benchmarking software Bench-MR [9], which allows us to demonstrate that our solution is superior to both state-of-the-art planners and the solution from [5] in the considered task of local maneuver generation.…”