UC Berkeley

As the emphasis on manufacturing is shifting from mass production to mass customization, the demands for flexible automation keep increasing. Human-robot collaboration (HRC), as an effective and efficient way to enhance flexibility, has attracted lots of attention both in industry and academia in the past decade. These robots are called co-robots. The fundamental research question is how to ensure that co-robots operate efficiently and safely with human partners.

To achieve that, two problems should be addressed: 1) co-robots should know the human's future trajectory and avoid potential collisions to guarantee safety, and 2) co-robots should know the human's intentions and take corresponding actions to ensure task efficiency. Therefore, a robotic system is adopted, which reasons about human behavior and makes human-aware planning using the reasoning information. For reasoning about human behavior, a hierarchical probabilistic modeling method and two online adaptation algorithms are proposed for human plan recognition and human trajectory prediction. The hierarchical probabilistic modeling method explicitly utilizes the hierarchical behavior of the human and uses a pipeline to identify human intention through the trajectory, the trajectory type, the action, and finally the plan, which is explainable and data-efficient.Two online adaptation algorithms, the adaptable neural network and the adaptable lognormal method, are proposed for short-term and long-term trajectory prediction. These two adaptation algorithms enable the prediction models to online accommodate different human behaviors and to deal with the lack of human data. For short-term prediction, the adaptable neural network utilizes the recursive least square-parameter adaptation algorithm to online adapt the last layer of a neural network model, the prediction of which is employed to the real-time collision avoidance. For long-term prediction, the adaptable lognormal method uses an objective-based adaptation algorithm to online adapt the sigma-lognormal model, from which the duration of the whole trajectory is estimated and later will be used in the task planner that optimizes the task completion time. For making human-aware planning, a separation task planner is proposed, which uses the knowledge of human intention and makes the robot execute actions that are parallel to the human actions, while minimizing the task completion time. Different from those methods that only consider time efficiency, this separation task planner additionally improves human satisfaction and avoids potential conflicts, which can be shown in the experiment results.