Our research is centered around two main lines:
First, we are using robots as models of natural cognitive processes. Recent progress in empirical sciences complemented by technological advances in robotics make it possible to instantiate complete scenarios in a robot and verify hypotheses about the mechanisms of natural cognition.
Second, we are investigating how an embodied and developmental perspective on perception and cognition applied to robots can push the state of the art in robotics. Robots will acquire greater autonomy and some hard problems (object recognition, for instance) can be turned into simpler ones by adopting an alternative view.
Quadruped robot Puppy
Our main robotic platform has been the compliant, underactuated quadruped robot Puppy, equipped with a rich sensory set which encompasses different sensory modalities (see Figure).
Path integration using multimodal proprioceptive sensory information
In this case study (Reinstein & Hoffmann, 2013), we have exploited the rich multimodal sensory set on the robot to develop a legged odometer. The odometer was used for aiding of an inertial navigation system using an Extended Kalman Filter, giving rise to a dead reckoning (path integration) system for the robot which was tested for different gaits and grounds.
Unsupervised terrain discrimination
In this ongoing case study, we are employing the same multimodal sensory set and investigating how different ground substrates on which the robot is running can be discriminated. We specifically focus on the respective influence of the different sensory modalities as well as on the effect of the action (the gait) used by the robot. Clustering is performed using a modification of the greedy Gaussian Mixture Models, a bio-inspired method related to the Neural Modeling Field Theory.