Real-World Data for DexArt
DexArt provides standardized evaluation for robot learning. Real-world data validates whether simulation performance transfers to physical hardware.
DexArt at a Glance
Benchmark Profile
DexArt is a benchmark for dexterous manipulation of articulated objects using multi-finger robot hands. Created by researchers at UC San Diego and Tsinghua University, it focuses on manipulating objects with joints and hinges — laptops, faucets, toilets, drawers — using simulated dexterous hands in SAPIEN.
The Sim-to-Real Gap
SAPIEN articulated object simulation simplifies joint friction and backlash. Simulated multi-finger contact misses real finger deformation and tactile feedback. Object material properties (plastic, metal, porcelain) affect manipulation but are simplified in simulation.
Real-World Data Needed
Real dexterous manipulation recordings of articulated objects with multi-finger contact data. Force/torque data during hinge and joint manipulation. Diverse articulated object instances with real material properties.
Complementary Claru Datasets
Custom Dexterous Manipulation Collection
Real multi-finger manipulation of articulated objects provides the contact dynamics simulation cannot faithfully model.
Force-Torque Manipulation Dataset
Force data during joint manipulation captures the resistance, friction, and compliance of real hinges and mechanisms.
Bridging the Gap: Technical Analysis
DexArt addresses a critical gap in dexterous manipulation benchmarks: most focus on grasping rigid objects, but real-world dexterous tasks often involve manipulating articulated objects — opening containers, turning handles, adjusting mechanisms. These tasks require coordinating multiple fingers to apply force at specific points while respecting the object's kinematic constraints.
The sim-to-real gap for dexterous manipulation of articulated objects is particularly wide. Real hinges have friction, backlash, and varying resistance that depend on material, lubrication, and wear. A simulated laptop hinge behaves identically every time; a real one varies across manufacturers, ages, and conditions. Multi-finger contact during these interactions involves complex finger deformation and slip dynamics that current physics engines approximate poorly.
Generalization across object instances is the benchmark's central challenge. A policy must learn to open laptops of different sizes, with different hinge stiffness, and different opening angles. Real-world data spanning diverse object instances provides the distributional coverage that simulated object randomization only approximates.
The benchmark's focus on everyday articulated objects connects directly to household robotics applications. A robot assistant that can open drawers, turn faucets, and operate appliance doors would address some of the most requested household tasks. Real-world data of humans manipulating these objects provides the demonstration data for learning these everyday skills.
Frequently Asked Questions
DexArt is a benchmark for dexterous manipulation of articulated objects using multi-finger robot hands. Created by researchers at UC San Diego and Tsinghua University, it focuses on manipulating objects with joints and hinges — laptops, faucets, toilets, drawers — using simulated dexterous hands in SAPIEN.
Real dexterous manipulation recordings of articulated objects with multi-finger contact data. Force/torque data during hinge and joint manipulation. Diverse articulated object instances with real material properties.
SAPIEN articulated object simulation simplifies joint friction and backlash. Simulated multi-finger contact misses real finger deformation and tactile feedback. Object material properties (plastic, metal, porcelain) affect manipulation but are simplified in simulation.
Yes. Claru coordinates data collection on specific robot platforms and in specific environments to enable direct comparison between simulated and real performance for DexArt tasks.
Get Real-World Data for DexArt
Discuss purpose-collected data to validate and improve your DexArt-trained policies on physical hardware.