blog
August 28, 2026

Authors:
Honglu He†, Jacob Laufer†, Zhiwu Zheng, David Elkan-Gonzalez, Raman Goyal, Xinyi Li, Su Lu, Mishek Musa, Berke Saat, Nicolas Tan, Colm Prendergast
†equal contribution
A complete loop from physical task design, through scalable demonstration collection, to a policy that fuses everything the robot can sense.
01 / Industrial Dexterity Benchmark | 02 / DAG-ROS | 03 / AG-iDP3 |
|---|---|---|
Benchmark boards | Imitation-learning stack | Multimodal diffusion policy |
Three low-cost, reproducible boards — datacenter fiber switches, automotive harnesses, and a planetary gearbox — built in the NIST Assembly Task Board tradition. | A ROS 2 framework tying together teleoperation, time-aligned multi-sensor capture, and real-time impedance control for rapid data collection and deployment. | A diffusion policy fusing wrist & scene RGB, point clouds, and proprioception with per-phase input gating, chained by a behavior tree at deployment. Adapted from Improved 3D Diffusion Policy (iDP3). |
Opinions expressed in the content posted here are the personal opinions of the original authors, and do not necessarily reflect those of Analog Devices, Inc (ADI) or its subsidiaries. The content is provided for informational purposes only and is not meant to be an endorsement or representation by ADI or any other party. This site may also provide links or references to non-ADI sites and resources. ADI makes no representations, warranties, or other commitments whatsoever about any non-ADI sites or third-party resources that may be referenced, accessible from, or linked to this site.
|
|---|
|
|
Each input stream is encoded and concatenated into a single observation, then a diffusion U-Net denoises it into a 15-step joint-space action chunk streamed to the controller.
Wrist RGB | Scene RGB | Point cloud | Robot state → | Diffusion Policy → | Joint actions |
|---|---|---|---|---|---|
|
|
|
|
|
|

A full walkthrough of the benchmark boards, the AG-iDP3 policy, and autonomous deployment on the datacenter board.
Six sensor configurations, 48 trials each, on the datacenter cable-cleaning task. Cleaning succeeded on 100% of trials in every configuration and is excluded from the total.
78%Best total success
| 36%Single-camera RGB | ~100Teleop demonstrations |
|---|
Configuration | Grasp | Insert | Total |
|---|---|---|---|
RGB Wrist + RGB Scene best | 94% | 62% | 78% |
RGB Wrist + Crosby ToF | 96% | 39% | 67% |
RGB + PC Wrist | 98% | 34% | 66% |
RGB Wrist + RealSense PC Scene | 88% | 38% | 60% |
PC Wrist only (egocentric) | 92% | 14% | 52% |
RGB Wrist only baseline | 48% | 52% | 36% |
The benchmark suite · real hardware
Each board is fully open — bill of materials, fabrication files, assembly instructions, and scoring rubric. Photos below are of the physical builds.
![]() | Datacenter cable manipulationDense connector spacing, constrained routing, and tight workspace access — replicating a populated rack switch. Two tasks (cable swap and fiber maintenance) across ethernet, SC, and LC connectors give six sub-tasks, with difficulty tuned by how many cables populate the board. |
![]() | ![]() |
Automotive cable harnessFlexible cables in constrained, partially occluded spaces. Route two 6-wire harnesses through clips and seat connectors into the correct receptacles — eight connectors at varying heights and orientations, with LED-verified pass/fail and an advanced through-wall routing option that stresses bimanual coordination.
| ![]() |
|---|---|
![]() | ![]() |
![]() | Planetary gearbox assemblyFine manipulation and sequenced assembly: mount the ring frame, insert the sun gear, seat three planet gears in mesh, cap the carrier, and torque the bolts in order. Scored on fastening, correct gear mesh, and whether the finished carrier rotates freely.
|
|---|---|
![]() | ![]() |
|
|---|