Toward Robots with Peripersonal Space Representation for Adaptive Behaviors

The abilities to adapt and act autonomously in an unstructured and human-oriented environment are necessarily vital for the next generation of robots, which aim to safely cooperate with humans. While this adaptability is natural and feasible for humans, it is still very complex and challenging for robots. Observations and findings from psychology and neuroscience in respect to the development of the human sensorimotor system can inform the development of novel approaches to adaptive robotics. Among these is the formation of the representation of space closely surrounding the body, the Peripersonal Space (PPS) , from multisensory sources like vision, hearing, touch and proprioception, which helps to facilitate human activities within their surroundings. Taking inspiration from the virtual safety margin formed by the PPS representation in humans, this thesis first constructs an equivalent model of the safety zone for each body part of the iCub humanoid robot. This PPS layer serves as a distributed collision predictor, which translates visually detected objects approaching a robot’s body parts (e.g., arm, hand) into the probabilities of a collision between those objects and body parts. This leads to adaptive avoidance behaviors in the robot via an optimization-based reactive controller. Notably, this visual reactive control pipeline can also seamlessly incorporate tactile input to guarantee safety in both pre- and post-collision phases in physical Human-Robot Interaction (pHRI). Concurrently, the controller is also able to take into account multiple targets (of manipulation reaching tasks) generated by a multiple Cartesian point planner. All components, namely the PPS, the multi-target motion planner (for manipulation reaching tasks), the reaching-with-avoidance controller and the humancentred visual perception, are combined harmoniously to form a hybrid control framework designed to provide safety for robots’ interactions in a cluttered environment shared with human partners. Later, motivated by the development of manipulation skills in infants, in which the multisensory integration is thought to play an important role, a learning framework is proposed to allow a robot to learn the processes of forming sensory representations, namely visuomotor and visuotactile, from their own motor activities in the environment. Both multisensory integration models are constructed with Deep Neural Networks (DNNs) in such a way that their outputs are represented in motor space to facilitate the robot’s subsequent actions.