Ensuring safety in physical human–robot interaction is challenging due to hardware and control architecture differences across robots, and is often implemented as system-dependent ad hoc approaches. To offer a holistic solution, we present a hardware-independent safety-focused admittance control approach, which promotes safety at the reference-generation level. This safety framework can restrict virtual dynamics through soft virtual bounds. Hard bounds are also introduced as a way to impose infinitely stiff soft bounds. As part of the overall approach, we also present a method for serial manipulator and multisegment entity collision avoidance by using partial Jacobians. In order to demonstrate the methodology's versatility across hardware platforms, we experimentally validate on two robotic systems: first, the Virtual Reality Exoskeleton (V-Rex), a nonanthropomorphic full-body haptic device composed of five robotic arms interacting with the body at the hands, feet, and pelvis; and second, the EXO-UL8, an anthropomorphic bimanual upper limb exoskeleton; which exist on opposite ends of the task/joint space control, nonredundant/redundant, off-the-shelf (industrial)/custom, nonanthropomorphic/anthropomorphic spectra. Experimental results validate virtual dynamics, soft bounds, hard bounds, and multi-arm collision avoidance on both systems. In all cases, both systems respect bound and collision constraints, supporting the approach as a safety-focused admittance control design.