Abstract

Co-manipulation requires two or more agents to coordinate through a shared object while exchanging information through motion and force. Although this capability is routine for humans, robotic systems still struggle to participate in such interaction naturally and effectively. This thesis addresses that gap by studying human co-manipulation behavior, developing methods to quantify performance, collecting a leader-follower dataset that isolates haptic communication, and training data-driven controllers from human demonstrations. First, cumulative and real-time metrics were developed to quantify human dyad performance. In particular, the behavioral modes "quickly" and "smoothly" were formulated as measurable performance objectives using completion time and squared jerk. Analysis of prior human-human data showed that high-performing teams adapt their behavior to the required task mode, and that high-performing "quick" teams exhibit larger and more dynamic interaction forces, supporting the hypothesis that interaction force serves a communicative role in haptic collaboration. Second, a virtual-reality-based leader-follower experiment was de-signed to isolate haptic communication. The experiment included leader-follower, leader-follower-follower, and leader-leader configurations across 18 tasks spanning isolated and combined degrees of freedom. This produced a synchronized dataset of force, motion, and survey data from 15 participant groups for analyzing human coordination and training learned controllers. Third, behavior cloning was applied to learn controller policies from the experimental data. Five neural network architectures were compared for both leader and follower prediction tasks. A GRU-based follower model achieved the best balance of accuracy and efficiency, with validation loss 0.0035 and normalized RMSE of 1-3%. In simulation, the learned follower reduced interaction forces relative to impedance-based baselines while maintaining task completion. In contrast, learned leaders performed substantially worse, indicating that behavior cloning is far better suited to the follower's reactive role than to the leader's goal-directed planning role. A hybrid controller combining learned and classical components provided a tunable trade-off between reliability and learned force-profile characteristics.Together, these results show that human co-manipulation perfor-mance can be quantified in ways useful for controller design, that interaction force likely plays an important communicative role, and that behavior cloning can capture important aspects of follower behavior from human demonstrations. The work establishes a foundation for future physical human-robot co-manipulation studies in which learned followers can be evaluated with human partners.

Degree

College and Department

Ira A. Fulton College of Engineering; Mechanical Engineering

Rights

https://lib.byu.edu/about/copyright/