Soft and Highly-Integrated Optical Fiber Bending Sensors for Proprioception in Multi-Material 3D Printed Fingers

TL;DR

Soft robots require distributed proprioception for reliable closed-loop control, but integrating low-cost optical sensing with flexible bodies is challenging. The authors present a semi-automated fused-filament fabrication process that embeds plastic optical fibers and flex-PCBs within TPU joints to create a monolithic bending transducer, with about 10 minutes of manual embedding. The resulting sensors achieve an average linearity of approximately $70\%$ and an RMS angular error of about $4.81^{\circ}$, while maintaining fingertip accuracy near $12\ \mathrm{mm}$ under external forces and consuming roughly $188\ \mathrm{mW}$ per finger at a cost near $10$ dollars per finger; a PCA-based approach enables online contact detection at $8\ \mathrm{Hz}$. Together, these results demonstrate a scalable, integrated distributed sensing solution for proprioception in soft, tendon-actuated fingers, highlighting potential for autonomous soft robotics and larger sensor networks.

Abstract

Accurate shape sensing, only achievable through distributed proprioception, is a key requirement for closed-loop control of soft robots. Low-cost power efficient optoelectronic sensors manufactured from flexible materials represent a natural choice as they can cope with the large deformations of soft robots without loss of performance. However, existing integration approaches are cumbersome and require manual steps and complex assembly. We propose a semi-automated printing process where plastic optical fibers are embedded with readout electronics in 3D printed flexures. The fibers become locked in place and the readout electronics remain optically coupled to them while the flexures undergo large bending deformations, creating a repeatable, monolithically manufactured bending transducer with only 10 minutes required in total for the manual embedding steps. We demonstrate the process by manufacturing multi-material 3D printed fingers and extensively evaluating the performance of each proprioceptive joint. The sensors achieve 70% linearity and 4.81° RMS error on average. Furthermore, the distributed architecture allows for maintaining an average fingertip position estimation accuracy of 12 mm in the presence of external static forces. To demonstrate the potential of the distributed sensor architecture in robotics applications, we build a data-driven model independent of actuation feedback to detect contact with objects in the environment.

Figures (8)

• Figure 1: (a) The multi-material 3D printed finger alongside a photo of one of the embedded flex-PCBs. (b) Each joint in the TPU flexure holds an embedded optical fiber bending sensor enabling light loss-based pose estimation for each joint.
• Figure 2: Schematic representation of the manufacturing and integration process of the sensorized multi-material finger. There are three main steps to the process: embedding of the optical fiber, embedding of the flex-PCBs, and printing of the finger. All the embedding steps are conducted during the printing process by temporarily pausing filament extrusion.
• Figure 3: Closeup photo of a pair of holders with an embedded optical fiber and a schematic detailed view of the embedding channel. The embedded fiber is able to support a mass of 300.
• Figure 4: Setup used for the characterization of the fingers. The photo is taken according to the POV of the camera used for the tracking of the markers.
• Figure 5: Characterizations of the sensors. (a),(b),(c) The plots show measured optical power loss against the MCP, PIP, and DIP joint rotations for each finger during the quasi-static trials with their respective best-fit lines. (d) The sensitivity achieved for each sensor is shown grouped by finger on the bar chart. (e) RMS error in degrees for each joint is shown. (f) Hysteresis as percentage of full scale range for each joint is shown.