From locomotion to cognition

In a series of case studies, we have demonstrated a developmental pathway in a quadrupedal robot Puppy.

1. In a first study [pdf], the robot first learned coordinated movement commands (gaits somewhat resembling those seen in nature – walk, bound, pace etc.) which later formed its motor repertoire.
2. In a second study [pdf], we let the robot explore different motor actions and recorded the corresponding sensory stimulations from its multimodal sensory set comprising primarily tactile and proprioceptive sensors. Then, we systematically analyzed the information flows between motors and sensors and showed how the robot could extract a primitive representation of its body – sensorimotor contingencies that are intrinsic to its morphology and nature of sensory apparatus – with minimal prior knowledge. We studied how these salient relationships change with different gaits as well as different ground substrates.
   1. From regularities in sensorimotor space to first goal-oriented behaviors. In subsequent case studies, we presented the robot with different scenarios, in which the robot needs to deploy the regularities it had learned in order to improve its performance in different tasks. In a studies on path integration [pdf1,pdf2], we devised a novel legged odometer design that combines the tactile and proprioceptive information into a stride length estimate in a model-free fashion (no explicit knowledge about the robot kinematics or dynamics is needed). Combined with an inertial navigation system (using an onboard inertial measurement unit), we obtained a dead reckoning architecture that proved to be robust to slippage and worked for different gaits and grounds. Other studies [pdf1,pdf2] addressed the recognition of different grounds and appropriate action selection: choosing the best gait in every context. Finally, another study [pdf], where a simplified wheeled robot was used, extended the robot’s niche to other agents: acting as “predator”, the robot was presented with a task of hunting its conspecific; successful solutions required to plan a sequence of actions as well as prediction of the “preys’s” movement.

In summary, the case studies constitute a coherent developmental pathway where the robot incrementally acquires simple models of its body and interaction with the environment and learns to use them to master tasks of increasing complexity.

The contributions to robotics are best summarized in this article [postprint-pdf] that provides an overview of how the underactuated nature of the robot Puppy on one hand speeds up learning and saves energy and, on the other hand, facilitates perception.

This work was the basis for my dissertation (defended in 2012 - pdf) and was conducted at the Artificial Intelligence Laboratory, University of Zurich, under the supervision of Prof. Rolf Pfeifer. It was funded by the SNSF “From locomotion to cognition” project and has continued under the EU project eSMCs “Extending sensorimotor contingencies to cognition”.