Adaptive Reachability Assessment in the Humanoid Robot iCub

We present a model for reachability assessment implemented in a simulated iCub humanoid robot. The robot uses a neural network both for estimating reachability and as a controller for the arm. During training, multi-modality information including vision and proprioception of the effector's length was provided, along with tactile and postural information. The task was to assess if a target in view was at reach range. After training with data from two different effector's lengths, the system generalised also for a third one, both for producing reaching postures and for assessing reachability. We present preliminary results that show good reachability predictions with a decrease in confidence that display a depth gradient.