Reproduce “Distilling Contact Planning for Fast Trajectory Optimization in Robot Air Hockey” on OpenResearch

About this paper

Julius Jankowski, Ante Marić, Puze Liu, Davide Tateo, Jan Peters, Sylvain Calinon

This work introduces a hybrid control framework for robot air hockey that combines a learning-based high-level contact planner with a model-predictive low-level motion controller. The approach enables agile robot behavior, demonstrating higher scoring accuracy and increased puck speed compared to existing purely control-based and learning-based methods in both simulated and real-world experiments.

Key insights

• Combining learning-based high-level contact planning with model-based low-level motion control effectively addresses the trade-off between optimality and real-time performance in dynamic contact tasks.
• Policy distillation using an Energy-Based Model enables real-time inference of optimal contact parameters, allowing for agile robot behaviors in contact-rich scenarios.
• The framework's tunability, through stochastic optimal control objectives and chance constraints, allows explicit shaping of robot behavior, prioritizing accuracy, speed, or a balance, and exploiting robot kinematics for complex maneuvers like bank shots.

Summary by alphaXiv.