TAN RUNJIAARMINE · ROBOTICS & EMBODIED AI
← ALL PROJECTS

DESAY SV SINGAPORE · MATRIXX

Cable Assembly & Contact-Guided Insertion.

Combining visual perception and contact feedback to make robotic connector insertion more reliable.

Collaboration
Desay SV Singapore
Challenge
Intrinsic AI for Industry Challenge
Milestone
MatrixX · Qualification Phase 7th

01 · THE CHALLENGE

Finding a port is only the first step

Cable assembly brings perception and physical interaction into the same task. A robot must locate the connector and its target, approach with the correct orientation, and respond when contact changes the motion. Small alignment errors can become failed insertions.

In collaboration with Desay SV Singapore, our team MatrixX explores how visual registration, contact feedback and learned skills can support this process. The work combines system integration with experiments in connector insertion.

02 · THE APPROACH

See, make contact, and adjust

  1. Locate and align

    Visual perception identifies task-board components and estimates the relative geometry needed to approach a connector and port.

  2. Respond to contact

    Contact feedback helps distinguish free motion from interaction with the socket, supporting alignment corrections during insertion.

  3. Develop reusable skills

    Policy-learning experiments investigate reliable single-end insertion, using demonstration data and feedback during development.

03 · PHASE 1 · JUL 2026

Two-end insertion record

The Phase 1 record follows a two-end cable-insertion task in simulation. The accompanying scene view shows the robot and task-board components used during development.

Phase 1 simulation record of the five-cable, two-end insertion task.
Phase 1 belief view of the robot, task board and cable-assembly components.
Phase 1 task-board scene in the belief view. Open the image to inspect the components.

04 · QUALIFICATION PHASE · 7TH PLACE

Automated HIL-SERL workflow

MatrixX placed 7th in the Qualification Phase. The workflow brings together visual preprocessing, automated data collection, policy learning and feedback during insertion.

Qualification Phase: simulated connector insertion with force and torque traces.
Qualification workflow: vision input, preprocessing, automated data collection, HIL-SERL policy learning, learned insertion policy and robot execution, connected by feedback.
Pipeline diagram from the July report. Its comparison with spiral search is qualitative; this slide does not provide a quantified benchmark.
Explore 7 qualification rollouts Different positions on two rails

These examples show individual insertion rollouts at the rail positions labeled in the source recordings.

Rail 0 · −0.060

Qualification Phase insertion rollout at this rail position.

Rail 0 · −0.030

Qualification Phase insertion rollout at this rail position.

Rail 0 · 0.000

Qualification Phase insertion rollout at this rail position.

Rail 0 · +0.030

Qualification Phase insertion rollout at this rail position.

Rail 0 · +0.055

Qualification Phase insertion rollout at this rail position.

Rail 1 · −0.050

Qualification Phase insertion rollout at this rail position.

Rail 1 · −0.030

Qualification Phase insertion rollout at this rail position.

05 · LABORATORY DEVELOPMENT

From perception to insertion

Selected examples from the 2026 project reports show the components under development. Each demonstration is labeled by its setup and stage.

REAL ROBOT · LABORATORY FIXTURE · JUN 2026

Insertion on a physical arm

A separate laboratory insertion experiment using a green fixture, shown alongside the HIL-SERL work in the June report. This demonstrates the physical test setup rather than a complete cable-assembly run.

HARDWARE PREPARATION · JUN 2026

SC connectors and NIC hardware

Optical cable with SC connectors and a network interface card prepared for experiments.
Physical optical connectors and a network interface card prepared for experiments. The photograph documents the hardware used to explore cable-assembly tasks.

06 · RESEARCH & ENGINEERING

Building reliability one skill at a time

Visual registration

Early iterations focus on recognizing components and connecting image observations with approach poses.

Contact-guided execution

Insertion experiments investigate contact response, alignment errors and the conditions under which a connector seats successfully.

Policy learning

Later reports explore single-end insertion policies and observer-guided learning as a route toward more reusable manipulation skills.

Simulation and physical experiments

The July videos document challenge development in simulation. The June fixture demonstration shows a separate physical setup. Each record keeps its stage and setup visible, while the team continues to develop reliable insertion across a wider range of conditions.

COLLABORATION

MatrixX · AI for Industry Challenge

Our team MatrixX placed 7th in the Qualification Phase of the Intrinsic AI for Industry Challenge. The project continues to connect research in robot learning with the practical demands of industrial assembly.

Official challenge page ↗