Real-World Data for PartNet-Mobility

PartNet-Mobility provides standardized evaluation for robot learning. Real-world data validates whether simulation performance transfers to physical hardware.

PartNet-Mobility at a Glance

Sim
Environment
Multi
Tasks
Standard
Evaluation
Active
Community

Benchmark Profile

PartNet-Mobility is a large-scale dataset and benchmark of articulated 3D objects with part-level annotations and mobility information. Built on SAPIEN, it provides over 2,000 articulated object models across 46 categories with annotated joints, motion ranges, and part semantics for training manipulation policies.

Task Set
Manipulation of articulated objects across 46 categories: cabinets, refrigerators, microwaves, laptops, scissors, and more. Each object has annotated joints (revolute, prismatic) with motion range limits.
Observation Space
Point clouds or RGB-D of articulated objects, joint state information (angles, positions), part segmentation masks.
Action Space
End-effector actions (grasp + pull/push) or joint-level torques applied to articulated object parts.
Evaluation Protocol
Manipulation success measured by achieving target joint states. Generalization across object instances within categories and across unseen categories.

The Sim-to-Real Gap

SAPIEN provides reasonable articulation simulation but simplifies joint friction, backlash, and wear. Real cabinet hinges, drawer slides, and laptop joints have resistance profiles that vary with age, load, and environmental conditions. Material interactions between gripper and object surfaces are simplified.

Real-World Data Needed

Real articulated object manipulation recordings with diverse object instances. Joint friction and resistance measurements from real objects. Multi-modal data (vision + force) during articulated object manipulation.

Complementary Claru Datasets

Custom Articulated Object Collection

Real manipulation recordings of diverse articulated objects capture the joint friction, resistance, and material properties simulation simplifies.

Force-Torque Manipulation Dataset

Force data during opening, closing, and adjusting articulated objects provides essential training signal for force-adaptive manipulation.

Egocentric Activity Dataset

First-person video of humans using articulated objects in daily life provides visual pretraining for household manipulation policies.

Bridging the Gap: Technical Analysis

PartNet-Mobility is the definitive resource for articulated object manipulation research, providing the 3D models and annotations that nearly all articulated manipulation work builds on. The benchmark's value lies in the scale and diversity of its object collection — over 2,000 models across 46 categories, each with annotated part hierarchy and joint definitions.

The sim-to-real gap for articulated objects is deceptively wide. In simulation, a cabinet door has a single revolute joint with constant friction. In reality, the same cabinet door has a hinge with varying friction depending on load and position, a soft-close mechanism that changes resistance near the closed position, and a latch or magnetic catch that creates a step change in required force. These joint dynamics vary across manufacturers, ages, and conditions in ways that parameterized simulation models cannot capture.

Generalization across object instances is the benchmark's central challenge and the area where real-world data has the highest impact. A policy must learn to open cabinets of different sizes with different hinge types, drawers with different slide mechanisms, and refrigerators with different handle configurations. Simulation can randomize parameters, but real-world data captures the actual distribution of joint characteristics that randomization only approximates.

The part-level annotation structure of PartNet-Mobility enables research on part-aware manipulation — policies that understand which part to grasp, which direction to apply force, and how the object's kinematics constrain the motion. Real-world data validating these part-aware policies ensures they work with the imprecise part segmentation that real perception provides, not just the perfect annotations available in simulation.

Key Papers

  1. [1]Xiang et al.. SAPIEN: A SimulAted Part-based Interactive ENvironment.” CVPR 2020, 2020. Link
  2. [2]Mo et al.. PartNet: A Large-Scale Benchmark for Fine-Grained and Hierarchical Part-Level 3D Object Understanding.” CVPR 2019, 2019. Link

Frequently Asked Questions

PartNet-Mobility is a large-scale dataset and benchmark of articulated 3D objects with part-level annotations and mobility information. Built on SAPIEN, it provides over 2,000 articulated object models across 46 categories with annotated joints, motion ranges, and part semantics for training manipulation policies.

Real articulated object manipulation recordings with diverse object instances. Joint friction and resistance measurements from real objects. Multi-modal data (vision + force) during articulated object manipulation.

SAPIEN provides reasonable articulation simulation but simplifies joint friction, backlash, and wear. Real cabinet hinges, drawer slides, and laptop joints have resistance profiles that vary with age, load, and environmental conditions. Material interactions between gripper and object surfaces are simplified.

Yes. Claru coordinates data collection on specific robot platforms and in specific environments to enable direct comparison between simulated and real performance for PartNet-Mobility tasks.

Get Real-World Data for PartNet-Mobility

Discuss purpose-collected data to validate and improve your PartNet-Mobility-trained policies on physical hardware.