Reproduce “Neural Collision Clearance Estimator for Batched Motion Planning” on OpenResearch

About this paper

J. Chase Kew, Brian Ichter, Maryam Bandari, Tsang-Wei Edward Lee, Aleksandra Faust

The research presents ClearanceNet, a neural network for efficient collision clearance estimation, and integrates it into CN-RRT, a sampling-based motion planner. This approach achieves up to an 845x speedup in batched collision checks, accelerates overall planning by up to 42%, and often yields paths up to 36% shorter than baselines, with real-world validation on a Fetch robot.

Key insights

• The minimum separation distance is a Lipschitz-continuous function of robot and workspace poses, making it effectively learnable by deep neural networks.
• Neural networks' inherent batch processing capabilities on GPUs can drastically accelerate collision checking compared to traditional geometric methods, which are sequential.
• The differentiability of a learned clearance model enables efficient gradient-based path repair, providing a more effective alternative to complete re-planning when minor collisions are detected.

Summary by alphaXiv.