Reproduce “Using Local Experiences for Global Motion Planning” on OpenResearch

About this paper

Constantinos Chamzas, Anshumali Shrivastava, Lydia E. Kavraki

The Global-Local Sampler (GL-Sampler) framework enhances sampling-based motion planners by decomposing workspaces into local primitives, learning efficient local sampling strategies, and synthesizing them into a global sampler. This approach consistently achieved solution times that were often orders of magnitude faster and significantly higher success rates compared to uniform sampling and CVAE methods in environments with narrow passages.

Key insights

• Decomposing a complex global planning problem into local, manageable primitives allows for effective transfer of learned sampling knowledge across diverse environments.
• Simple geometric local primitives, even when exhibiting decomposition or similarity errors, can robustly guide sampling in high-dimensional and complex global configuration spaces.
• Gaussian Mixture Models (GMMs) provide an efficient and interpretable way to store and synthesize local sampling strategies, avoiding the need for complex, environment-specific learning models.

Summary by alphaXiv.