Refine ForceTwin dynamics online from robot interaction

Develop and evaluate a method for refining a ForceTwin physics-informed digital twin online using interaction forces collected by a robot, particularly when changes such as loading a drawer or readjusting a door closer alter the instance-specific dynamics.

Background

ForceTwin performs one-shot, per-instance identification from instrumented human probing. The identified dynamics can become inaccurate when the physical state of the object changes, such as when a drawer is loaded differently or a door closer is readjusted.

The paper explicitly identifies online refinement from robot-generated interaction forces as unresolved. Solving this problem would allow a deployed robot to update the digital twin as it encounters changing object dynamics, extending the static identification procedure to adaptive manipulation.