Inverse Detection for Deterministic Safety Monitoring in Dynamic Industrial Environments

The rapid growth of e-commerce, warehouse automation, and flexible production increases the need for robotic systems that can operate safely and efficiently in shared workspaces. In such environments, human operators, industrial robots, and automated vehicles increasingly work in close proximity. Conventional safeguarding based on physical separation or protective stops when a monitored region is entered can limit both flexibility and productivity. Collaborative robots offer an alternative, but their practical implementation often still relies on worst-case safety assumptions. This dissertation investigates the limitations of existing safety concepts for human-robot collaboration and develops deterministic approaches to improve selectivity and productivity without abandoning the fail-safe logic required in industrial safety systems. The work focuses primarily on two safety principles: Power and Force Limiting (PFL) and Speed and Separation Monitoring (SSM). First, the reliability of controller-provided external force and torque estimation in collaborative robots is examined. Experimental measurements on a commercially available collaborative robot show that controller-provided external force estimates should not be interpreted as direct force measurements. The estimated forces are strongly configuration-dependent and are affected by model uncertainty, friction, dynamic effects, sensor noise, and proximity to singularities. As a result, PFL applications require conservative thresholds and operating limits. This restricts the applicability of PFL in tasks involving higher speeds, larger payloads, or more dynamic robot motion. Second, SSM is analyzed as a pre-contact safety strategy. SSM allows the robot speed to be adapted according to the separation distance between the human and the robot. Although safety standards permit dynamic and adaptive separation distances, industrial implementations often rely on static or discretely switched protective fields based on worst-case values. Important limitations include the lack of safety-rated information about human velocity and motion direction, restricted access to validated robot and task information, constraints in safety-rated communication, and limited mechanisms for continuously adapting protective fields. The analysis therefore also investigates how robot position, task state, stopping behavior, rail motion, and task-relevant reach can be used to reduce unnecessarily large monitored regions when this information is available with the required safety integrity. To address the selectivity limitation of conventional presence detection, the concept of inverse detection is introduced. Instead of directly attempting to classify humans, inverse detection starts from a conservative default rule: every detection is treated as human-equivalent unless it can be associated with a deterministically verified non-human object. Only detected objects for which the geometry, pose, state, communication status, and sensor measurements satisfy predefined verification conditions are filtered from the safety evaluation. All unknown, inconsistent, or unverified objects remain active and can trigger a speed reduction or protective stop. This principle shifts the focus from probabilistic human classification to deterministic verification of known non-human objects. A proof-of-concept implementation was developed and evaluated in an industrial logistics testbed. The results show that robot-manipulated objects and communicating mobile objects can be selectively filtered under predefined verification conditions. The conservative default response is reactivated when communication is lost, geometric consistency fails, a timeout expires, or verification is incomplete. The implementation demonstrates the technical feasibility of the inverse-detection concept. A central conclusion of this dissertation is that an important limitation of current collaborative safety systems lies not only in the availability of sensing technologies, but also in the lack of deterministic mechanisms for selectively interpreting sensor data in a way that can be validated for safety-related use. By combining a conservative default response with deterministic verification and explicit uncertainty margins, this dissertation identifies a path toward more selective and productive safety monitoring while retaining the fundamental principles of industrial safety.