EmbodAI

State-of-the-art (deep) reinforcement learning systems, for all their fantastic achievements, struggle in real-world tasks that are trivial for humans, especially those involving physical interactions. At the same time these systems consume excessive power for training and operation. That is because they are inefficient with their model representations (many parameters) and their data (big data and many trials for training). We see the (robot) body as an enormous computational resource that is poorly understood and largely underappreciated. But how to harness embodiment remains an important open question.

Signal processing by a physical body is extremely cheap and robust, but specialized; the brain is flexible, but more power-hungry. This results in a design trade-off that leads us to the following two multidisciplinary research questions for this project.

1. Which continuous-learning processing tasks should be delegated primarily to hardware (body) and which primarily to software (brain), and

2. how should the brain and the body be designed to capitalize on the potential of embodied intelligence?

We study how we can harness Embodiment as a resource in a next generation of EAI systems.

Supervisory team:

• Youri van de Burgt, Associate Professor, Microsystems, ME, Promotor
• Irene Kuling, Assistant Professor, Robotics, ME, co-promotor
• Thijs van de Laar, Assistant Professor, BIASlab, EE, co-promotor