Open-ended learning: a test
One of the most promising ideas in robotics, on which the PILLAR project is based, is open-ended learning: instead of being programmed for a fixed task, a robot explores its environment, discovers on its own what it can do in it, and builds a repertoire of skills over time, much like a child at play. The catch is that pure curiosity has no direction. A robot left free to explore may become skilled at many things, but not necessarily at the things its user actually needs.
This is the problem at the heart of the PILLAR project: how do we keep the flexibility of autonomous, curiosity-driven learning while gently steering it towards a purpose — the mission (purpose) assigned to the robot by its user?
The answer developed in WP3 is the Motivational Engine, the component that manages the robot’s different “desires” and decides, moment by moment, what is worth doing: pursuing the assigned task, practising skills it has not yet mastered, or exploring to discover new possibilities.
The key mechanism is what we call the prototype: an abstract representation of the features that goals fulfilling the purpose tend to have in common. The prototype can be learned by the robot through its own experience or communicated directly — in one line of research, simply by describing the purpose in natural language, with a Large Language Model helping the robot identify which objects in the scene are relevant to it. Either way, the prototype acts as a bias in the robot’s autonomous learning: the robot remains free and curious, but among the many things it could learn, it prioritises those that “look like” its purpose.
Crucially, this is no longer just a design on paper. The Motivational Engine was implemented in an extended version of the H-GRAIL architecture and tested in two settings: a simulated environment (with both independent and interdependent goals, and with the task changing across domains) and a real robotic setup.
In both cases, the comparison was the same: a robot whose autonomous learning is biased by the purpose versus an identical robot without this mechanism. The results are consistent: the purpose-aware robot learns the skills relevant to the user’s request more quickly and satisfies the request more reliably across changing domains, without sacrificing its ability to explore and learn beyond the task itself.
The Motivational Engine is now operational and is being integrated into the overall cognitive architecture of PILLAR-Robots (WP5) — a step towards lifelong-learning robots that are autonomous and useful.