Collision Detection (Robot Safety)

Detecting, in real time while the robot moves, that it has hit a person or object, and immediately stopping or yielding.

Collision detection, in this safety sense, is a real-time robot function: the moment the body unexpectedly hits a person or object, the controller must notice as fast as possible and respond, such as stopping, backing off, or switching into a compliant mode. Most collaborative robots don't rely on skin-like sensors, using only proprioception instead: a dynamics model predicts the joint torque normal motion should require, that prediction is compared against the actual torque measured from motor current or joint torque sensors, and a difference beyond some threshold is flagged as a collision. The generalized momentum observer used by De Luca, Haddadin, and colleagues is the classic method; their 2017 survey in IEEE Transactions on Robotics breaks the process into detection, isolation (which link was hit), and identification (how large the impact force was). A manufacturer's 'collision sensitivity' setting is exactly this threshold: set it too low and the robot stops on false alarms; set it too high and a real collision hits harder before it's caught. Franka's libfranka, for example, lets each joint have separate 'contact' and 'collision' thresholds, and the robot stops moving once the collision threshold is exceeded.

ExampleA collaborative arm's elbow bumps into a nearby worker's shoulder while carrying something; the joint's measured torque spikes above the model's prediction and crosses the threshold, triggering a stop — the worker feels only a gentle nudge — after which the error has to be cleared before operation can resume.

Also called

Collision Protection, Sensorless Collision Detection

Related

Generalized Momentum Observer · Collision Reaction · Power and Force Limiting · Protective Stop · Joint Torque Sensor · ISO/TS 15066 Robots and Robotic Devices — Collaborative Robots

Sources

Haddadin, De Luca, Albu-Schäffer, Robot Collisions: A Survey on Detection, Isolation, and Identification (IEEE T-RO 2017)
libfranka robot.h（setCollisionBehavior / automaticErrorRecovery）