When two humans perform a collaborative manipulation task, they leverage an intuitive understanding of which motions are natural and safe for their interaction partner. Intuition lets human collaborators predict both the feasibility of an action, as well as their partner's ergonomic preference for one feasible action over another. This mutual understanding in human-human teams allows for comfortable and efficient collaboration. However, in human-robot teams, robots typically lack the models which would give them this same understanding of human action choice. This problem is compounded by that fact that humans are unique. A model that accurately predicts the actions of one human collaborator may be incorrect for another.

This dissertation explores how personalized mechanical models and ergonomic cost functions for humans can endow collaborative robots with an understanding of human action feasibility and ergonomic preference analogous to that possessed by human teammates. Specifically, we focus on the task of robot-to-human object handoffs. We show that planning handoffs with knowledge of a human's arm kinematics yields improved ergonomics and safety, and is preferred by human collaborators. Next, we demonstrate how ergonomic cost functions which predict a human's preference for different feasible actions can allow co-robots to shape human action choices, helping humans choose globally ergonomically optimal actions in multi-part tasks. Finally, we show how these ergonomic cost functions can be learned online, allowing co-robots to quickly adapt to the preferences of an individual human partner.