Scalable Open-Source Visuotactile Sensor for 6-Axis Contact Wrench Estimation in Tensegrity Robots

Technical Summary: Scalable Open-Source Visuotactile Sensor for 6-Axis Contact Wrench Estimation in Tensegrity Robots

Problem Statement
Tensegrity robots, characterized by networks of rigid struts and tensile cables, offer high strength-to-weight ratios and resilience, making them suitable for exploration and search-and-rescue. However, their high degrees of freedom and nonlinear dynamics complicate state estimation, which is critical for feedback control and locomotion. While existing proprioceptive methods rely on strain sensors, cable encoders, or IMUs, they lack reliable direct ground contact measurement. Current contact sensing solutions for tensegrity robots often suffer from drift, hysteresis, or fabrication complexity. Furthermore, while visuotactile sensing has advanced in robotic manipulation and legged locomotion, scaling these sensors for large, meter-scale tensegrity structures remains challenging due to fabrication difficulties, the need for adhesives that may fail under impact, and the lack of robust, open-source designs.

Methodology
The authors propose a scalable, open-source visuotactile endcap sensor designed specifically for tensegrity robots. The system comprises three primary mechanical components:

1. Elastomer Shell: A compliant shell made of Smooth-Sil™950 (Shore 50A) that forms the primary contact surface. It features an inner surface with shallow dents painted with permanent marker dots to serve as visual markers for deformation tracking.
2. TPU Interface: A 3D-printed thermoplastic polyurethane (TPU) layer that connects the elastomer to the rigid base.
3. Rigid Base: Houses a miniature USB camera with a 180-degree fisheye lens and a custom LED ring for uniform illumination.

A key innovation is the gyroid-infill bonding technique. Instead of using adhesives, the 3D-printed TPU interface is printed with a continuous-infill gyroid pattern. During the casting of the elastomer, liquid silicone flows into these open infills, creating a durable mechanical interlock upon curing. This eliminates the need for adhesives, which often fail under shear or tensile stress.

For data processing, the system employs a tactile-to-wrench neural network. The input is a dense, two-channel shear vector field ($V \in \mathbb{R}^{2 \times H \times W}$) derived from optical flow analysis of the camera images (using methods from GelSlim 4.0). The output is a six-dimensional contact wrench ($w = [F_x, F_y, F_z, T_x, T_y, T_z]^\top$). The model utilizes a deep residual Multi-Layer Perceptron (MLP) trained on a dataset of 34,368 samples collected using a KUKA Med R820 robot arm and an ATI Gamma force/torque sensor.

Key Contributions

• Scalable Sensor Design: A 130 mm diameter endcap sensor that matches the scale of meter-level tensegrity robots, utilizing off-the-shelf components and low-cost 3D printing.
• Adhesive-Free Bonding: A novel mechanical bonding method using gyroid infill to join elastomer and TPU, resulting in a stronger and more consistent interface than traditional adhesives (Sil-Poxy™), as verified by tensile and peel testing.
• Open-Source Implementation: The design, CAD files, and code are made publicly available to facilitate replication and integration into various tensegrity platforms.
• 6-Axis Wrench Estimation: A lightweight neural network capable of mapping shear vector fields to full 6D contact wrenches, enabling both force estimation and binary contact detection.

Experimental Results

• Bonding Strength: Tensile and peel tests demonstrated that the mechanical gyroid bonding provides approximately 25% higher tensile strength and 300% higher shear strength compared to Sil-Poxy™ adhesive, with lower variation.
• Static Validation: On a held-out validation set, the model achieved a Mean Squared Error (MSE) of 0.1531. The model accurately reproduced dominant force directions and magnitudes, with force estimation proving more robust than torque estimation.
• Dynamic Validation: Under dynamic motion (circular trajectories), the MSE increased to 2.67. While torque prediction showed higher variability and phase lag due to camera frame rates and processing latency, the system successfully tracked main temporal force trends.
• System Integration: The sensor array was integrated onto a 12 kg, three-bar tensegrity robot. In tests involving suspended, single-contact, two-contact, and three-contact configurations, the system reliably identified ground contacts, with force estimates for contacting endcaps rising sharply while non-contacting endcaps remained near zero.

Significance and Claims
The paper claims that this system substantially improves the practicality of tactile feedback for tensegrity robots by offering a low-cost, reproducible, and physically interpretable pathway toward contact-aware proprioception. By providing explicit constraints on which endcaps are anchored to the ground, the sensor transforms floating-base estimation problems into more constrained formulations. The authors position this work as a step toward robust autonomous operation in complex environments, noting that while torque estimation requires further refinement (e.g., via temporal modeling), the system already provides reliable binary contact cues and useful force estimation for state estimation and control. The open-source nature of the project is highlighted as a means to promote broader adoption and reproducibility in compliant robotic systems.