New Publication on Integration and Continual Learning-Based Modeling of a Soft Robotic Sensor
KTH’s Robot Design Lab and Honda Research Institute Japan contribute on embedded soft sensing for social robots, accepted for presentation at ICRA 2026 in Vienna, Austria.
Reliable proprioception remains a central challenge in soft robotics, where compliance and safety often come at the expense of accurate state estimation. In this work, we present the integration of a multi-material 3D-printed soft strain sensor directly into the soft continuum actuator forming the neck of HARU, our tabletop social robot for child-robot interaction. By co-extruding conductive and non-conductive thermoplastic polyurethane in a single fabrication step, we embed sensing functionality within the actuator body itself, preserving its external geometry and compliance while enabling bidirectional angle estimation.
Several conductive path geometries were evaluated, with a gauge-type configuration selected for its linearity and repeatability. Four sensors were arranged in a cross-configuration within the actuator, enabling differential measurements and improved robustness compared to prior standalone soft sensor demonstrations. To translate resistance signals into angle estimates, we compared linear regression, a static neural network, and a continual-learning model with online parameter updates. Across structured and randomized motions, as well as repeated experimental sessions, the continual-learning approach consistently achieved the highest accuracy.
This work advances scalable and practical proprioceptive sensing strategies for soft robots by combining embedded multi-material fabrication with adaptive learning. The integration of robust sensing into HARU’s compliant neck contributes toward safer, more expressive, and reliable interactive robotic platforms suitable for real-world social environments.