Catch, Throw, Repeat: Planning for Human–Robot Partner Juggling

How can a robotic system be designed to perform dynamic, contact-rich tasks like partner juggling alongside a human? The authors investigate the challenges of real-time coordination, predictive motion planning, and robustness to human variability in a shared three-ball cascade juggling task.

The authors propose a control architecture that integrates three primary components: a predictive ball-tracking pipeline using Kalman filters, an adaptive online trajectory optimization method, and a state-machine-based coordination logic. The trajectory planner uses a multiple-shooting formulation to generate smooth, dynamically feasible motions that adapt to the human's throws in real-time. The system was tested on a 4-DOF robotic manipulator using a funnel-shaped end-effector, with performance evaluated through a user study involving eight participants of varying skill levels.

The system demonstrated high reliability across all participants. In the three-ball cascade setting, every participant exceeded the previously reported best-case performance of four consecutive robot catches, with some achieving up to 20. In single-ball trials, the system achieved a 100% success rate over 40 consecutive cycles. The study highlights that while success rates naturally decrease as the number of balls increases due to tighter timing constraints, the architecture remains robust enough to handle the variability inherent in human throwing.

This work advances the field of physical human-robot interaction by moving beyond static handovers to dynamic, high-speed collaboration. By demonstrating that a robot can adapt to the unpredictable timing and trajectories of a human partner in a complex task like juggling, the authors provide a scalable framework for future applications in shared autonomy and dynamic manipulation.