Conditioned Learning of Sampling Distributions for Efficient Manipulator Motion Planning

Kurzfassung

Motion Planning for Manipulation faces complex challenges due to high-DoF robots, complex environments, tasks, and frequently additional constraints. This complexity can reflect in long, undesirable computation times. To speed up planning, Machine learning (ML) has been extensively researched, with many hybrid methods leveraging widely adapted Sampling-Based Motion Planning (SBMP) algorithms. Repetition Sampling (RS) previously employed prior experiences to model task-relevant portions of the configuration space by a Gaussian Mixture Model, which can then replace the Uniform sampling strategy within SBMP. However, while most robotic tasks display repetitiveness, intra-task variability is common due to changing task contexts, such as varying object positions or start configurations. Depending on the context, queries that appear similar from a high-level perspective (e.g., grasping an object) can result in solution paths composed of distinct configurations. \par For this purpose, we propose \textit{Conditioned Repetition Sampling} (CRS), which extends on RS and enables faster planning by narrowing sampled regions down further to context-relevant configurations. We show how to train a GMM representing the joint probability distribution of context information and configurations and how to efficiently condition it on the query context during planning using Gaussian Mixture Regression (GMR), a method successful in Learning from Demonstration. We validate our approach in three increasingly complex experiments in simulation and on the mobile manipulator AIMM of the German Aerospace Center (DLR). We also investigate Constrained Motion Planning, which is more rarely focused on by other methods. We show that CRS can bias SBMP's exploration towards configurations that likely contribute to finding a solution, resulting in significant speed-ups of between 1.3x and 3.7x compared to RS and up to 13.7x compared to standard Uniform Sampling. Overall, our analysis shows that CRS reliably finds motion plans while offering the flexibility to condition on various context information, such as workspace poses and end effector orientations. Additionally, CRS offers increased interpretability of the planning pipeline compared to related methods using Deep Learning, making it an overall promising option for efficient Motion Planning.