Human-Informed Robot Agility: Understanding Human Pushing Interactions for Skill Transfer to Humanoids
Cornelia Bauer and Nancy S. Pollard

Abstract

This research focuses on the analysis of human dynamic interactions and their applicability to robot systems. Studying human mechanics and control strategies in these interactions provides valuable insights for developing control approaches for dynamic interactions in robots. This work captures and analyzes wall push-offs as a specific interaction, and develops a skill model capturing push-off behaviors adaptable to different robot embodiments. Pushing skills are modeled as position trajectories relative to the center of mass,  implicitly encoding interaction forces. We show that this approach can be scaled and adapted through optimization, and present results for transferring wall push-offs to two different robot platforms.

[paper] 

Transfer of human push-off skills to robots

Based on human push-off demonstrations, we model push-offs as pushing skill trajectories with respect to the center of mass (CoM) frame of the human body. We can transfer these trajectories to a robot as follows: We first copy them into the CoM frame of the robot, then transform the trajectories into the shoulder frame, and scale them to lie within the robot arm workspace. To account for different dynamic properties between robots and humans, we introduce skill parameters that allow us to further modify the push, and find suitable values through optimization. A more detailed description of this approach can be found in our paper.

We test our approach on two robots: The CMU Ballbot, a human-sized robot balancing on a single spherical ball, and the PushBot, a small omnidirectional robot platform which uses a hovercraft-like lift system.

Wall Push-Off MoCap Dataset

MoCap Mass Model

Starting from the skeleton model of the MoCap data format, we assign a mass to every bone.  Each bone is modeled as a cylinder, where the length equals the length of the MoCap model bone segment length. We assume a uniform density of 1  gram per cubic centimeter across all body segments. This mass model is scaled to individual study participant body weight estimates. For each bone, we can further compute an inertia tensor corresponding to a cylinder, and combine all bone inertias to compute the overall body inertia, which changes shape depending on body pose. The resulting cylindrical bone model is shown here.

Body Momentum time series

The following graphs visualize momentum time series (x -  component, y - component, and magnitude) over time during wall push-offs for different subjects. The image title contains entry angle, speed, and exit angle. Hand contact with the wall is indicated in red.