Interaction of the robotic limb with the vulnerable environment requires a task-dependent trade-off between compliance (to avoid damaging organic matter) and rigidness (to exert the required forces on the environment for gripping and manipulating organic matter). Such trade-offs may be realised by inherent direction-dependent material properties (adhesive gripping, sticking), by low-level control over limb shape (enabled by additive manufacturing or smart mechanical design), low-level adaptations in limb impedance (enabled by soft variable impedance actuators and or admittance control), or high-level control by a (trained) human operator. In case of robot companions or assistants in domestic environments, soft robotic limbs should be able to be guided or gently pushed away like with a pet. A challenging type of interaction occurs when a (trained) human operator is required to directly control soft robotic limbs, or – where needed – to teach, correct, or supervise (semi-)intelligent limbs. What interfaces need to be designed to allow effective control and feedback to the operator? Flexible, multi-segmented soft robotic limbs have many more degrees of freedom and many more possibilities for adapting compliance, all of which poses formidable challenges in terms of sharing or trading control, (haptic) feedback to the operator, and situation awareness. The integrated, distributed neuromuscular control of animals such as cuttlefish will serve as inspiration to enable intuitive hierarchical control over soft robotic limbs, e.g. during soft-robotic assisted surgery.