Learning to Plan via Neural Exploration-Exploitation Trees

Chen, Binghong, Dai, Bo, Song, Le

arXiv.org Machine Learning 

Planning paths efficiently in a high-dimensional continuous state and action space is a fundamental yet challenging problem in many real-world applications, such as robot manipulation and autonomous driving. Since the general path planning problem is PSPACE-complete (Reif, 1979), one typically resorts to approximate or heuristic algorithms. Sampling-based planning algorithms, such as probabilistic roadmaps (PRM) (Kavraki et al., 1996), rapidlyexploring random trees (RRT) (LaValle, 1998), and their variants (Karaman & Frazzoli, 2011), provide principled approximate solutions to a wide spectrum of high-dimensional path planning tasks. However, these generic algorithms typically employ a uniform proposal distribution for sampling which does not make use of the structures of the problem at hand and thus may require lots of samples to obtain an initial feasible solution path for complicated tasks, e.g., a narrow passage in a map. To improve the sample efficiency, researchers designed algorithms to take problem structures into account, such as the Gaussian sampler (Boor et al., 1999), the bridge test (Hsu et al., 2003), the reachability-guided sampler (Shkolnik et al., 2009), the

Duplicate Docs Excel Report

None found

Similar Docs  Excel Report  more

None found