Abstract
Reproducibility in robotic manipulation research is often limited by the physical availability of specific objects across different laboratories. CapriDB addresses this challenge by providing a low-cost, end-to-end pipeline for capturing, printing, and tracking objects. By combining handheld photogrammetry with 3D printing, researchers can create "digital twins" that are easily shared and physically replicated. This paper details the pipeline's workflow and introduces a database of 40 high-quality textured meshes to facilitate standardized benchmarking in the field.
Introduction
In robotics, reproducing experiments involving object manipulation is notoriously difficult. A study conducted in one part of the world using local consumer products is hard to replicate elsewhere if those exact items are unavailable. While initiatives like the YCB Object and Model Set provide a physical baseline, they rely on shipping physical kits, which is slow and expensive. Furthermore, fixed "off-the-shelf" objects prevent researchers from systematically varying physical properties like mass distribution or friction. CapriDB offers a decentralized alternative: a data-centric approach where objects are shared as digital files and manufactured locally.
The CapriDB Pipeline
The authors propose a three-stage workflow: Capture, Print, and Innovate.
Capture and Post-processing
The pipeline begins with handheld photogrammetry. Instead of expensive industrial scanners, a standard camera is used to take approximately 40 photos of an object from various angles. To assist the reconstruction software in orienting camera positions, objects are placed on high-texture backgrounds like newspapers.
The resulting raw meshes undergo post-processing using open-source tools. MeshLab is utilized to remove background elements, while Blender is used to fill holes, refine the geometry, and scale the model to precise metric units.
Print and Track
The refined 3D models are sent to 3D printing services. This stage allows for significant innovation: researchers can print replicas in different materials to test friction sensitivity or modify internal structures to create custom mass distributions.
To close the loop, the pipeline integrates with the SimTrack visual tracking system. The high-resolution 3D models allow for real-time 6D pose estimation using standard RGB or RGB-D cameras, ensuring that the robot's perception matches the physical reality of the printed object.
The CapriDB Dataset
The paper introduces the CapriDB database, consisting of 40 watertight, textured mesh models. The collection spans several categories, including office supplies, kitchen items, tools, and toys. Notably, it includes digital models of objects from the 2015 IEEE ICRA Amazon Picking Challenge. Each entry provides the Wavefront OBJ mesh, texture maps, physical dimensions, and the original source photos, offering a comprehensive resource for the research community.
Experimental Validation
To validate the pipeline, the authors conducted grasping experiments using a KUKA industrial arm equipped with a Schunk SDH dexterous hand. The experiments demonstrated that 3D-printed replicas provide sufficient visual and geometric fidelity for robust tracking and manipulation.
The robot successfully executed various grasps based on poses estimated from the CapriDB models. The tracking system proved resilient to occlusions caused by the robot's hand, confirming the utility of the high-quality digital models in complex, real-world scenarios.
Conclusion
CapriDB shifts the paradigm of manipulation research from hardware-centric to data-centric. By enabling the local manufacture of standardized or custom-modified objects, it lowers the barrier to entry for high-quality robotics research and significantly improves experimental reproducibility. Future work will focus on automating the post-processing stages and exploring the "grasp moduli space" to understand how subtle shape deformations impact robotic grasping success.
Summary updated