Co-Learning of Task and Sensor Placement for Soft Robotics

Unlike rigid robots which operate with compact degrees of freedom, soft must reason about an infinite dimensional state space. Mapping this continuum space presents significant challenges, especially when working a finite set discrete sensors. Reconstructing the robot's from these sparse inputs is challenging, since sensor location has profound downstream impact on richness learned models for robotic tasks. In work, we present novel representation co-learning placement and complex Specifically, neural architecture processes on-board information to learn salient selection placements optimal task performance. We evaluate our model learning algorithm six robot morphologies various supervised tasks, including tactile sensing proprioception. also highlight applications motion subspace visualization control. Our method demonstrates superior performance in algorithmic human baselines while latent spaces that are semantically meaningful.