Automated Air Hockey Inspired Dual Arm Hitting Framework

The idea

Learning how an object moves after a hit needs a lot of consistent impact data - tedious and hard to gather by hand. This project automates it with a dual-arm setup that hits an object back and forth, like a collaborative game of air hockey, continuously generating labelled impact examples.

Method

• Automated data collection: two arms rally an object between them; a finite-state machine coordinates the exchange so hits are repeatable and self-resetting.
• Impact-aware Extended Kalman Filter: approximates the discontinuous impulse dynamics through a hitting-force model that balances the energies during collision, keeping the automation robust to impacts.
• Full density modeling: friction stochasticity and control error make the outcome uncertain, so the relationship between hitting impulse and the object's resulting displacement is modelled as a full probability distribution rather than a point estimate.

Application - the "Golf" principle

The learned motion model is used to plan sequential hits with two or more robots: much like sinking a long putt in several strokes, a chain of planned hits can drive an object to a location far beyond the reach of any single robot.