Robots are capable to autonomously handle more and more complex tasks. However, today's robots either have their workspaces physically separated from the humans' ones or their abilities severely restricted when acting near humans. In this thesis, we investigate several approaches allowing to plan not only for the robot but also for the human, enabling to predict and elicit their decision process and actions, leading to better human-robot interactions (HRI).First, we show through a user study, why such a scheme applied to robot navigation is crucial for efficient and satisfactory interaction. Planning for the robot and the human allows to find solutions in intricate situations where collaboration is necessary but also for the robot to be proactive and legible while navigating. Secondly, we alleviate from inherent ephemeral nature of interaction in collaborative navigation to explore how co-planning can be applied for task planning. We introduce a new referring expression generation algorithm, using an ontology as a knowledge base, and show that it is the fastest one to date while being designed for HRI application. We use it in a human-aware task planner to estimate the feasibility and cost of communication during task planning, preventing deadlock or suboptimal plans. Finally, a novel approach to human-aware task planning is proposed where action models and decision streams of the robot and the human are distinct and used to produce conditional plans.