Building dynamics models through contact discontinuitites

This demo demonstrates learning object mesh and fricition parameters of a cube by observing its contact rich interactions with the flat table. I worked with Penn PhD student Mathew Halm, Penn PhD student Bibit Bianchini and Penn faculty Michael Posa.

The setup is a Franka robot arm that is tasked to toss a cube that is tracked with three point grey cameras. The contactnets model learns the object mesh that is presented via an interactive browser. We see a good prediction of the object mesh in around 20 tosses of the cube.

The thumbnail and the video used in this page is taken with permission from Bibit Bianchini’s project page, which further details the implementation of this work.