The Impact of Approximate Methods on Local Learning in Motion Planning

Diane Uwacu, Chinwe Ekenna, Shawna Thomas and Nancy Amato
Machine learning methods have been applied to many motion planning algorithms including probabilistic roadmap methods (PRM). There are many variants of these methods and choosing the best one every time is hard and depends on local properties of the environment. A successful learning approach has been developed to offset this issue. This learning approach was applied to PRMs to help decide intelligently what method to utilize in dynamically created local regions of the environment or task space. It used exact neighbor finding approaches and removed the need to partition environments to get improved results.
In this work we make further advances by introducing approximate neighbor finder methods. It has been established that approximate neighbor finding methods are faster than exact methods, still work well in connecting nodes to edges in PRMs, and that connection is robust to noise. We study what happens when noise is introduced into learning by using approximate methods instead of already studied exact methods. We show that the impact of noise on learning depends on how much learning needs to take place given the topology of the environment. Our results demonstrate a correlation between heterogeneity and the need for learning over a local region.