Chopstick-Inspired Non-Prehensile Manipulation
for Datacenter Automation

Harshit Khurana, Elliott Hogg, Freddie Hong, Iason Sarantopoulos, Fabian Otto, Iris Andrussow, Hugh Williams, Tim Deegan, David Sweeney, Andromachi Chatzieleftheriou

Microsoft Research, Cambridge, UK

Dense cable environment behind a datacenter switch
Dense cable clutter at the back of a datacenter network switch.
Human operator parting cables to reach a port
A human parts cables to access a transceiver port — the behaviour we automate.

Abstract

Automating repair and maintenance tasks in datacenters requires robots to operate in dense cable environments where overlapping cables occlude target components and must be displaced without damage. Traditional prehensile approaches are impractical here due to the difficulty of isolating individual cables and the risk of entanglement. We present a non-prehensile cable parting approach using a chopstick-inspired tool that slides between cables and creates localized clearance around a target transceiver port. The tool and motion primitives are designed to enable safe cable displacement without explicit force sensing or cable-specific planning. We evaluate the approach across randomized cable configurations in both a MuJoCo physics simulation and real-world experiments using a dual-arm UR5e system, where one arm declutters cables and a second arm inserts a transceiver into the cleared port. Successful end-to-end transceiver insertion is achieved in the integrated multi-robot pipeline.

Key contributions

01

Chopstick tool

A tapered two-prong tool that gently parts dense, fragile cables without grasping them.

02

Motion primitives

Four modular primitives — Insert, Rotate, Comb, Open — composed into interpretable policies.

03

Real-world validation

Sim + dual-arm UR5e experiments reaching 100% clearance and end-to-end transceiver insertion.

The chopstick tool

Rather than relying on precise end-effector force sensing — which becomes intractable with many simultaneous cable contacts — we simplify the interaction through tool design. The end-effector consists of two parallel prongs actuated by a servo hinge, with tapered tips to slide between packed cables, a reach tuned to enter dense bundles without wrist collisions, and a parallel opening that separates cables evenly along its length.

Traditional chopsticks inspiration
(a) Inspiration
Chopstick tool closed
(b) Closed
Chopstick tool open
(c) Open

Motion primitives

Instead of planning over continuous cable state, manipulation is decomposed into four modular primitives. Every policy begins with Insert and ends with Open; the intermediate primitives vary.

Insert primitive
Insert — slide closed chopsticks into the bundle.
Comb primitive
Comb — translate laterally to separate cables.
Rotate primitive
Rotate — orient the subsequent open.
Open primitive
Open — spread the prongs to create clearance.

Policies

P1 Insert → Open
P2 Insert → Rotate → Open
P3 Insert → Comb → Open
P4 Insert → Comb → Rotate → Open

The Comb primitive is the critical differentiator: by laterally separating overlapping cables before opening, policies P3 and P4 stay robust even in chaotic configurations.

Experimental setup

A dual-arm system with two Universal Robots UR5e manipulators on a shared base: one carries the chopstick tool for decluttering, the other a Robotiq 2F-85 gripper for insertion, in front of an ARISTA 7050QX-32S switch populated with QSFP transceivers and fiber optic cables.

Dual-arm UR5e experimental setup
Dual-arm UR5e platform: chopstick arm (left), gripper arm with RealSense D405 (right).

Environment classes

Structured cabling
Structured
Semi-chaotic cabling
Semi-Chaotic
Chaotic cabling
Chaotic

Cable parting: before & after

Segmentation before parting
Before — cables occlude the target port.
Segmentation after parting
After — clearance region is cleared.

Results

Standalone decluttering clearance rate (%)

PolicyStructuredSemi-ChaoticChaoticAll
P19380057.7
P27367749.0
P31001009397.7
P4100100100100

Policies without Comb (P1, P2) collapse in chaotic clutter, while P3 and P4 achieve near-perfect clearance across all environment classes.

Clearance rate and occlusion reduction plots
Clearance rate and occlusion reduction across policies and environments.
Sensitivity to occlusion threshold
Clearance vs. occlusion threshold — P3/P4 remain stable across thresholds.

Full task: declutter + insert

PolicyParting clearanceInsertion successFull task success
P493.3%78.6%73.3%
Insertion success by scenario
Insertion success by scenario group.

Attempting insertion directly into cluttered ports succeeds only 6.7% of the time. With chopstick decluttering (P4), the integrated pipeline reaches 73.3% end-to-end success — a >10× improvement — recovering most of the insertion capability lost to cable clutter.

Video

Overview of the chopstick cable-parting approach and end-to-end transceiver insertion on the dual-arm UR5e platform.

BibTeX

@misc{khurana2026chopstick,
  title  = {Chopstick-Inspired Non-Prehensile Manipulation for Datacenter Automation},
  author = {Khurana, Harshit and Hogg, Elliott and Hong, Freddie and
            Sarantopoulos, Iason and Otto, Fabian and Andrussow, Iris and
            Williams, Hugh and Deegan, Tim and Sweeney, David and
            Chatzieleftheriou, Andromachi},
  year   = {2026}
}