Physical human–robot interaction (pHRI) is becoming essential in several domains, such as industry and medicine. The robustness, repeatability, and precision of robots are no longer sufficient to meet the growing complexity of required tasks. These tasks also demand human supervision to leverage human adaptability and situational assessment. Furthermore, performing certain physical tasks can lead to musculoskeletal disorders (MSDs). Consequently, pHRI—and in particular physical human–robot collaboration (pHRC)—is being actively investigated to enhance performance in motor tasks while improving the health and safety of human operators. The doctoral work presented here is part of this broader research effort. Physical collaboration between a human and a robot is a topic that researchers approach in various ways. One approach, still relatively underexplored, consists of first studying collaboration between two humans—that is, placing human behavior at the center of robotics research. Indeed, pHRC is the most natural and comfortable form of collaboration for humans, as evidenced by subjects who report preferring a human partner over a robotic one. The objective of this thesis was therefore to identify specific features and strategies emerging during human–human collaboration and to transfer them to human–robot collaboration. To this end, we defined a collaborative pick-and-place task, with variable parameters such as object mass, and conducted the experiment with 30 dyads (pairs of subjects). The results demonstrated substantial inter-dyad variability in behavior during the task, which can complicate the modeling of human behavior, and revealed that some of the variable parameters had no measurable effect on task performance. These findings enabled the construction of a human behavior model using inverse optimal control (IOC) methods. The generalizability of the resulting model was evaluated and showed that its use for trajectory prediction was well justified. The prediction method developed consisted of leveraging past behavior to estimate future behavior within a sliding temporal window. The long-term goal of this approach is to make the robot partner more proactive, capable of anticipating human motion, and thereby improving human comfort and safety during collaboration.