Smart Compliant Robots

In our research we focus on the question how much intelligence can be distributed over the robot's body, brain and world in order to achieve more computation ­efficient and human­ environment proof robots that have advanced cognition skills.

Current trends indicate that compliant actuators and structural elements are the key for a safer, more robust and energy efficient human­-robot interaction. Due to the increased use of compliant elements, the focus on the control algorithms has been shifted towards both machine learning techniques for modelling and control, and embodiment. By techniques based on neural networks and reinforcement learning such as deep reinforcement learning, the robot can learn a model-free adaptive control framework through imitation or interactions. On the other hand, by embodiment we explore the possibility to (partially) outsource functionality and computational tasks to compliant body elements of the robot which drastically reduces the complexity of the motor control.

Robots and other embedded systems often come with limited computational power. Today, the focus often lies on advanced computationally demanding techniques, which often run in the cloud. In order for those systems to become more independent of internet access or to reduce the required bandwidth for data communication, we explore the portability of machine learning techniques to embedded computing resources. We explore the bounds this imposes on different performance measures and the ways complex tasks and machine learning techniques can best be partitioned to optimally exploit local resources. While all the above involves fundamental research, we also focus on transferring our results to small and medium-sized companies.


Tony Belpaeme, Joni Dambre, Francis wyffels.


Jeroen Burms, Jonas Degrave, Luthffi Ismail, Gabriel Urbain, Brecht Willems.


Key publications