YCB Object and Model Set

• YCB Object and Model Set is a benchmark platform featuring 77 standardized household objects with precise physical measurements and detailed digital scans.
• It integrates high-resolution RGB–D data and geometric models to enable replicable evaluation of manipulation algorithms on platforms like ROS, OpenRAVE, and Gazebo.
• The set supports cross-system comparison through standardized scoring metrics and task protocols for applications in robotics and prosthetics research.

The Yale–CMU–Berkeley (YCB) Object and Model Set is a rigorously curated physical and digital collection designed to standardize benchmarking in robotic manipulation, prosthetics, and rehabilitation research. The set consists of 77 meticulously measured household objects and tools, complemented by high-resolution RGB–D scans, physical properties, and geometric models. These resources facilitate replicable, quantitative evaluation of manipulation algorithms, planners, control strategies, and mechanical design within a unified experimental framework, supporting integration with common software platforms such as ROS/MoveIt, OpenRAVE, and Gazebo (Calli et al., 2015).

1. Object Catalogue and Physical Properties

The YCB set is organized into five categories: Food Items (19 objects), Kitchenware (14), Tools & Hardware (20), Shapes & Miscellaneous (15), and Standard Manipulation Tests (3 setups plus timer). For each object, the mass (measured via digital scale, ±1 g), principal dimensions (via calipers, ±0.1 mm), nominal material, and rigidity are documented. No static or dynamic friction coefficients have been published as of the reference date.

Category	Representative Objects	Example Mass/Dimensions/Material
Food Items (19)	Coffee can, cereal box, apple, sugar box	Coffee can: 510 g, 80×195×80 mm, steel, rigid
Kitchenware (14)	Plate, bowl, mug, frying pan, pitcher, sponge	Plate: 279 g, 258×10×258 mm, ceramic, rigid
Tools & Hardware (20)	Hammer, screwdriver, wrench, scissors	Hammer: 688 g, 32×280×40 mm, steel/wood, rigid
Shapes & Miscellaneous (15)	Spheres (marble–softball), foam bricks, rope	Marble S: 3.6 g, Ø24.7 mm, glass, rigid
Manipulation Tests (3+1)	Box and blocks kit, 9-peg-hole, assembly toys	Box and blocks: 100×1" cubes, 2 containers

Material and rigidity are as supplied (e.g., solid lacquered wood for fruit, consumer-grade steel for tools, compliant foam for sponges and some shapes). All dimensions correspond to those most relevant for grasping and manipulation. The object set also includes flexible items (rope, chain) and semi-rigid items (cardboard boxes, PVC card) (Calli et al., 2015).

2. Model Acquisition and Digital Representations

RGB–D and geometric data acquisition utilizes the BigBIRD scanning rig, incorporating five calibrated RGB–D sensors and five high-resolution RGB cameras. Each object is scanned over 120 turntable increments (3° steps), yielding 600 RGB–D point clouds and 600 high-resolution RGB frames per object. Intrinsic and extrinsic calibration data as well as binary foreground masks (from Poisson mesh reprojection) are provided.

3D meshes are constructed by Poisson surface reconstruction, producing watertight triangle meshes. Texturing is achieved by UV-mapping per-vertex color from high-resolution images, and mesh accuracy preserves sub-millimeter surface details (holes remain in unsampled regions such as transparent or highly reflective surfaces). The data packages include OBJ (geometry + UV), MTL (materials), PLY (colored–point clouds), and STL (optional mesh) files, with PNG/JPG texture maps. Each mesh is aligned to a global coordinate frame (Z-up, X out toward camera, Y right) with the origin at the turntable center and units in meters (Calli et al., 2015).

3. Benchmark Scoring Formulas

YCB benchmarking protocols use a collection of standardized scoring metrics, primarily based on success/failure counts and weighted sums. All major formulae are explicitly defined:

• Pitcher–Mug Pour Score: For a single trial,

$$s = \frac{M_\mathrm{after} - M_\mathrm{before}}{P_\mathrm{before} - P_\mathrm{after}}$$

where $M_{\text{before/after}}$, $P_{\text{before/after}}$ are pre- and post-pour mug and pitcher masses. Score is credited only if at least half the mug is filled.

• Table-Setting Score:

$$S = 4N_\mathrm{in} + 2N_\mathrm{touch} + 1N_\mathrm{lifted\,only}$$

Quantifies correct zone placement versus partial or incorrect placement.

• Block Pick-&-Place Score:

$$S = 4N_\mathrm{within} + 2N_\mathrm{touching} - N_\mathrm{off}$$

Uses analogously defined placement criteria.

• Peg-Insertion Success Rate:

$$\mathrm{SuccessRate}(\delta) = \frac{N_\mathrm{inserted}}{N}$$

For $N$ trials with perturbation magnitude $\delta$.

All scoring is supported by task-specific constraints and initializations standardized across laboratories, enabling quantitative cross-system comparison (Calli et al., 2015).

4. Benchmarking Protocols and Task Templates

Each protocol applies the physical set and a five-part specification (task description, setup, robot, procedure, constraints). Principal protocols include:

• Pitcher–Mug Task: Pouring rice or water from pitcher to mug across ten predefined layouts, scored as above.
• Gripper Assessment: Fixed home grasp on each object and tool, tested under systematic $±1$ cm X/Y/Z offsets; scored per offset as 0–2 points.
• Table Setting: Pickup and placement of tableware into color-coded zones; simulatable in Gazebo or MoveIt.
• Block Pick & Place: Randomized arrangement and sequential placement of wooden blocks onto a template; measures open-loop grasping performance.
• Peg Insertion Learning: Linear-Gaussian controller for inserting pegs under varying board perturbations ($±5$ mm, $±10$ mm), scored by success rate.
• Box and Blocks (Prosthetics Test): Maximizing block transfers in two minutes with a heuristic grasp strategy.

All tasks provide both physical and simulation (URDF, KinBody) setup files; example baseline results are reported, such as HERB robot achieving 0.98 ± 0.00 pour score on 8/10 pitcher–mug trials (Calli et al., 2015).

5. Integration with Manipulation and Simulation Frameworks

Digital models and task scripts are compatible with leading robotics middleware. For ROS/MoveIt, each object is distributed as a URDF link referencing textured and collision meshes. Scene launch files and Setup Assistant configurations are provided, supporting planners such as OMPL, CHOMP, and CBiRRT with task-space regions (TSRs).

OpenRAVE integration is enabled via KinBody XML descriptors pointing to the mesh files, with sample scripts for instantiation. In Gazebo, a complete “ycb_world.urdf” spawns the full set on a table, and demonstration plugins (e.g., for pick-and-place) are available.

Watertight meshes can be preprocessed with FCL or QSDF libraries to generate signed-distance fields for fast collision checking in optimization-based planners. Mesh decimation utilities (e.g., Meshlab, VTK) facilitate tradeoff studies between planning speed and model fidelity (Calli et al., 2015).

6. Applications and Significance in Manipulation Research

The YCB Object and Model Set delivers a well-characterized and reproducible platform for benchmarking a wide spectrum of robotic manipulation methods. Its diversity in object properties (size, shape, material, rigidity), combined with high-fidelity digital assets and standardized protocols, enables consistent evaluation across domains: planning, learning, mechanical and gripper design, perception, and control.

By providing baseline experimental results and task templates, the set fosters comparability, transparency, and replicability in academic and industrial research. The inclusion of widely accepted manipulation tests and integration with standard simulation tools further cements its utility as a reference for both algorithmic development and hardware benchmarking (Calli et al., 2015).

7. Limitations and Future Extensions

As of the referenced release, the set does not specify static or dynamic friction coefficients for objects, though these are planned for future extensions. Some object meshes exhibit holes where transparent or highly reflective surfaces impede scanning; these regions are intended for improvement via photometric capture in subsequent releases. The collection can be acquired via research tutorials or at cost, subject to ongoing updates in both physical and digital resources (Calli et al., 2015).