Robotics Benchmark on Transfer Learning: a Human-Robot Collaboration Use Case

In the context of human-robot collaboration (HRC), the model of the robots needs to be adapted to describe new tasks in new environments and under new operating conditions. A long-standing challenge in HRC is to transfer the acquired robot's skills and adapt models from a limited amount of data and/or with limited computational resources. To facilitate the research addressing this issue, this paper proposes a transfer learning benchmark in a HRC setting using data acquired from a 7-DOF Franka Emika Panda robot. The goal is to estimate a dynamical model mapping set-point pose, measured pose, and velocity of the end-effector into the external interaction wrench. This type of problem may arise in real applications to design virtual sensors for forces/torques. In the proposed benchmark, the model can be estimated based on a long dataset acquired under a nominal operating condition of the robot, along with 5 shorter trajectories (to be used for model adaptation) gathered for 5 different values of translational stiffness. Performance is measured on 5 test trajectories where the same translational stiffness values of the transfer experiments are applied. Baseline results are presented using a transfer learning approach tailored to dynamical systems recently proposed in the literature.