In Work Package 4, our goal is to enhance the adaptability of robots in various situations for learning and task execution. In open environments, adaptive behaviours hinge on adjusting perception and action.
In open environments, adapting behaviours relies on adapting perception and action. Our aim is to enhance the autonomy of these adaptations. For instance, when the objects surrounding a robot change in two different contexts, even if its primary purpose remains constant (such as manipulating objects), it requires training to perceive, plan, and act in diverse ways.
To achieve this, we are developing world models that prioritize the crucial information obtained by robots from their sensors. These models predict the outcomes of specific actions or the utilization of particular skills. We want to give robots the ability to learn on their own. When reward signals are absent, we establish a curriculum of aptitudes and program robots to acquire fundamental and essential skills, such as navigating, controlling their arms, or grasping objects in various ways. We design algorithms that enable robots to explore their surroundings safely and discover new interactions. If the exploration becomes too complex, and an expert user is available to provide written instructions or a demonstration, we facilitate the robots’ learning from that source.
Lastly, when skills start to be mastered, we robustify and distill them into more general skills, then compose them to try to find new skills and test the limits of the current models and representations. Ideally, the robots should perform what is called “open-ended learning”: as long as the environments to which they are confronted present new challenges, they should keep learning and their skills should keep getting better. Even with relatively well-defined purposes, we know that practising in a controlled environment is not the same as the real world, so we are figuring out ways to make sure robots can adapt and do well in all sorts of situations and constraints.