Projecting Dancing Movements From Human Onto Robot Arm by Weighted Multipoint Mapping

Nowadays, robot choreography is a powerful medium for entertainment and product showcasing. Many artists perform alongside robots, both anthropomorphic and otherwise, to enhance their shows, and an increasing number of robotics companies use dance to highlight their products' capabilities. However, creating robot choreographies is a time-intensive and complex endeavour for both artists and engineers. Inspired by learning from the demonstration approach to plan robot movements, we have investigated novel strategies that use the human body as a teacher for manipulators.In these studies, we observe that audiences tend to seek a human-like quality in robotic movement, focusing on the coherence of the robot's full-body motion rather than the precise positioning of its end-effector. Furthermore, the movement of each body part is equally significant for the overall choreography. For these reasons, we propose a strategy in this context to dynamically control and weight multiple points along the robot's kinematic chain, using recorded data. To achieve this, we apply a reinforcement learning approach, and we test the proposed method in a specific use case: controlling both the robot's end-effector pose and its center of gravity on predefined trajectories.