Source note
Industrial Dexterity Benchmark: A Hardware-Software Benchmarking Platform for Industrial Dexterous Manipulation
Summary
The paper introduces the Industrial Dexterity Benchmark (IDB), a hardware platform for testing industrial manipulation, together with imitation-learning infrastructure and a multimodal diffusion policy. On the datacenter cable-cleaning task evaluated here, the best sensor configuration achieved 78% combined grasp-and-insert success using about 100 teleoperated demonstrations per task phase.
Problem
- Industrial tasks such as cable routing, connector insertion, and precision assembly remain difficult to automate because they combine tight clearances, deformable objects, occlusion, and contact-rich control.
- The problem matters because datacenter maintenance operates in densely populated racks with uptime targets above 99.99%, while classical vision-and-control pipelines can be brittle and costly to recalibrate for small task changes.
Approach
- The authors design three low-cost IDB boards for datacenter cable manipulation, automotive cable harnesses, and gearbox assembly; this paper reports experiments only on the datacenter board.
- DAG-ROS provides ROS2-based teleoperation, time-aligned sensor collection, real-time control, dataset processing, and deployment infrastructure.
- AG-iDP3 combines RGB features from R3M, point-cloud features from PointNet, joint positions, and wrist force/torque data in a diffusion U-Net that predicts 15-action chunks; successive chunks are temporally ensembled into 50 Hz commands.
- A modality-gating mechanism supports sensor ablations and phase-specific inputs. Wrist wrench data was retained for insertion, where contact information helped, and gated off for grasp and cleaning.
Results
- On the IDB Board #1 cable-cleaning task, evaluated with 48 trials per configuration, the best multimodal expansion Diffusion Policy reached 78% combined grasp-and-insert success.
- The 78% result exceeded the single-camera RGB Diffusion Policy baseline, which achieved 36% success, an absolute improvement of 42 percentage points.
- Each tested policy required approximately 100 teleoperated demonstrations per task phase.
- The evaluated model variants ranged from 68.8 million parameters for point-cloud-only iDP3 to 114.9 million for the two-camera RGB Diffusion Policy; the combined point-cloud-plus-RGB model used 94.6 million parameters.
- The reported evidence covers one task on IDB Board #1; the automotive cable-harness and gearbox boards are proposed benchmarks whose performance is reserved for future work.