Millimeter cable-driven continuum robots exhibit shape conforming, dexterous manipulation capabilities in constrained environments. They are increasingly used for narrow space and endoluminal intervention. For delicate manipulation, quantifying the force interaction between the robot and its surrounding environment is important for both shape adjustment and avoiding damages to luminal structures. In this work, we propose a real-time, whole-body contact estimation framework for small-scale continuum robots, based on actuation fibers and model-informed neural networks. The physical relationship among external body contact, internal actuation, and shape sensing of the continuum robot is formulated based on rod theory, and body contact estimation is treated as an inverse problem given the actuation tension profile and robot shape as inputs. The contact position and force are estimated using a neural network, and a generative adversarial network-based data augmentation strategy is proposed to reduce the need for large amounts of real data from the continuum robot under external forces. In addition, an automatic data acquisition platform is developed to efficiently collect the small amount of required data. Experiments with notched continuum robots were conducted to demonstrate the general applicability and accuracy of the proposed approach. The results show that the mean estimation errors for the three-dimensional (3D) contact position and contact force magnitude are 1.7 mm (2.3%) and 8.7 mN (5.8%), respectively, with an estimation frequency of 25 Hz. It paves the way for embodied integration using multiplexed fibers for the simultaneous actuation and sensing of millimeter-scale continuum robots, enabling their safer operation in confined spaces through machine intelligence.
Keywords: body contact; contact estimation; continuum robot; fiber multiplexing; robot manipulation.