Abstract
Traditional methods for designing custom robotic manipulators are often manual, time-intensive, and prone to failure when modifying complex CAD constraints. Furthermore, integrating tactile sensing is typically treated as a secondary process, leading to bulky or poorly fitted sensor arrays. This paper presents an end-to-end design pipeline that leverages graph grammars for topological synthesis and cage-based deformation for geometric optimization. By automating the generation of manufacturing files for both 3D-printed structures and custom-knitted tactile gloves, the system enables the creation of functional, sensor-enabled manipulators in minutes.
Introduction
In soft and rigid robotics, the ability to tailor a manipulator to a specific task—such as handling delicate objects or operating tools—is critical. However, the current design workflow is often a "trial-and-error" process. Engineers frequently spend weeks in CAD software, where minor adjustments to finger length or joint count can break parametric constraints, necessitating a manual rebuild of the entire assembly.
Moreover, the "sense of touch" remains difficult to integrate. Most tactile sensors are rigid and bulky, making them incompatible with the complex, organic geometries required for advanced manipulation. This research addresses these challenges by introducing a unified, compiler-like workflow that translates high-level design intent into physical, tactile-enabled hardware.
The Design Pipeline
The proposed pipeline bridges the gap between digital design and physical fabrication through a multi-stage process that decouples topology from geometry.
1. Topological Synthesis via Graph Grammars
The system utilizes a set of formal grammar rules to assemble manipulators from a library of modular components, including palms, knuckles, and phalanges. This ensures that every generated design is topologically valid and that all components maintain correct connectivity. This approach allows designers to explore a massive design space—estimated at $10^{49}$ possible configurations—without the fragility inherent in traditional CAD systems.
2. Geometric Optimization via Cage-Based Deformation
Once the basic structure is established, users can modify the geometry using a low-resolution "cage" mesh. This allows for intuitive stretching, widening, or tapering of parts. To prevent deformation from compromising mechanical integrity (e.g., distorting a circular joint pin), the system employs function-oriented part division. Mechanical joints are treated as "protected" zones that only scale uniformly, while the shafts of the phalanges can be freely manipulated to meet task-specific requirements.
3. Automated Tactile Mapping
Rather than embedding rigid sensors, the pipeline treats sensing as a conformal, knitted textile. Users select specific faces on the digital model for tactile sensitivity. The software then automatically generates a quad-dominant "stitch mesh," which is used to produce a 3D knitting pattern for an industrial knitting machine.
Manufacturing and Fabrication
The pipeline outputs two distinct sets of manufacturing files:
- STL Files: Used for 3D printing the rigid components of the hand using micro-carbon fiber filled nylon.
- DAT Files: Used for industrial knitting machines to create a two-layer tactile glove. This glove is embedded with piezoresistive fibers that detect pressure changes across the surface of the manipulator.
Experimental Results
The versatility of the pipeline was validated through the creation of four task-specific prototypes:
- The Egg Picker: Designed with forked fingers and wide tips to securely cradle and shake an egg without causing damage.
- The Wing Screw Driver: Features a single, offset rigid finger capable of 360-degree rotation without colliding with the screw's base.
- The Bottle Sorter: Utilizes tactile feedback to distinguish between full and empty water bottles by "weighing" them, allowing the system to decide whether to pour out the contents or discard the container.
- The Scissor User: A three-finger design capable of donning office scissors. The system uses pressure spikes in the tactile glove to detect when a cut has been successfully completed.
Quantitatively, the pipeline reduces the design phase from days to minutes, generating all necessary fabrication files automatically once the user finalizes the digital model.
Limitations and Future Work
While the pipeline significantly streamlines design, several challenges remain. The knitted sensors are currently susceptible to noise and environmental drift, requiring manual calibration before use. Additionally, while design and fabrication are automated, the final assembly—including cable routing and stitching the glove—remains a manual process. Future iterations could address these issues by integrating Reinforcement Learning (RL) to automatically optimize hand shapes for complex tasks and improving the durability of the drive cables, which are currently prone to breakage under high loads.
Conclusion
This integrated pipeline represents a significant step toward the automated co-design of robotic hardware and sensing. By combining graph grammars with flexible geometric deformation and automated textile manufacturing, the system allows for the rapid deployment of custom, tactile-aware manipulators tailored to specific real-world applications.
Summary updated