Centralize or Distribute: Integrating Action and Perception for Soft Robotic Manipulation

How do different systems perceive and respond to their environments? This fundamental question drives research in sensory-motor control, particularly in the contrast between centralized and distributed strategies. Biological systems offer compelling examples, particularly in case of soft embodiments such as elephant trunks and climbing plants. Animals rely on a central nervous system to process sensory information, predict outcomes and plan movements, whereas plants adapt through distributed control, responding locally to stimuli without a central processor. While both exploit tactile sensing and integrate multiple modalities for effective interaction tasks, they differ significantly in their integration process. Inspired by these biological models, this thesis explores how centralized and decentralized control paradigms can enhance soft robotic manipulation. Centralized approaches exploit the perceptual representation of the body and the environment by the central nervous system for action planning. Our investigation has enriched soft robots with innovative blind object recognition using a soft gripper and a multimodal perception building strategy for soft bodies through generative models. In contrast, distributed strategies rely on initiating and adapting preprogrammed behaviors without centralized perception. This approach has inspired innovative behavioral control mechanisms for robotic systems, particularly in reaching and contact tasks. Applying these strategies to soft robots explores how distributed sensing can trigger reactive actions, how soft-bodied agents manage action selection, and how lifelong-reaching mechanisms can enhance adaptability in dynamic environments. Through a comparative analysis, this work examines the strengths and limitations of each paradigm for soft robot applications concerning sensory integration, adaptability, computational efficiency, and the use of internal models. The findings reveal that neither paradigm is universally superior; rather, the choice depends on the specific control requirements and the performance demands of the application. Finally, the thesis highlights promising future directions, including advanced perception-building methods, lifelong learning frameworks, and adaptive control strategies, aiming to facilitate the real-world deployment of soft robotic systems.