This technology is a robot system and learning data generation method that determines collisions in real-time by generating training data from the variance between control target values and actual measured values of robot joints during non-collision states, and predicting dynamic normal operating ranges using an AI learning model.
Conventional torque sensor-based collision detection involves high hardware costs, motor current-based methods are prone to false positives due to friction, and existing AI approaches often suffer from reduced robot durability during the collection of actual collision data.
This technology proposes a method that calculates time-series maximum and minimum measured values from normal, non-collision operation data using sliding window and moving average techniques, utilizing them as training data to predict dynamic collision ranges. This enables accurate collision detection without the need for additional sensors. It can be applied to safety certification for collaborative robots and industrial manipulators, replacing expensive torque sensors while ensuring both safety and cost-efficiency.
This invention was developed with support from the Ministry of Trade, Industry and Energy for the development of deep reinforcement learning-based collaborative task technology capable of intelligently responding to unstructured work environments, such as assembly tasks.
N/A
