Identifying Physical Interactions in Contact-Based Robot Manipulation for Learning from Demonstration

Identifying physical interactions between a robot and its environment can improve autonomous robot manipulation. Methods in the literature can identify meaningful physical interactions, such as constraints, from a single task demonstration. However, they do not scale well when a wide variety of different physical interactions may be encountered. To alleviate this limitation, this work proposes to model physical interactions with interaction frames: reference frames that attach to the robot and/or ground body and have interaction classes associated with each axis. Interaction frames are identified by minimizing and decoupling the Cartesian mechanical power components in demonstration data. Thereby, interaction frames identify what interactions take place where, for example at geometric features. The method is evaluated in three experiments. First, identification accuracy is evaluated through several single-constraint experiments comparing velocity-based, force-based, and the proposed power-based method, where the latter is found to be advantageous. Second, the method is applied to a task demonstration that contains different sequential contacts. Third, it is illustrated how the method can be used as a basis for Learning from Demonstration, by reproducing the sequential task from a single demonstration. The method can be applied with little prior information, to advance the development of versatile robots.