A Reliable Learning-Based Approach for Self-Collision Checking on a Redundant Manipulator
Hua Liu, Hengsheng Wang, Xinping Guo, Liang Wang · IEEE Transactions on Industrial Informatics · 2025
In many real-world deployments, manipulators require efficient self-collision detection. Recent research has utilized NN-based detection methods to accelerate collision detection by directly learning the relationship between the collision state and joint values. However, certain inputs can cause the NN to produce high-probability but incorrect predictions, and this unreliability hinders the deployment of such methods in safety-critical domains. To address these challenges, first, we proposed a learnable feature embedding method to improve model accuracy (ACC). Second, we developed a monitor to identify these inputs by testing the NN model. The ACC, runtime, and reliability of our approach are assessed by conducting experiments on a redundant manipulator. The results demonstrate that the method achieves a 97% success rate and only requires 65% of the time compared to the state-of-the-art geometric approach. Furthermore, the monitor improves the detection rate of erroneous inputs by 33%.