Embodied AI & RoboticsSingapore & SEA10 weeks pilot, 8 weeks rollout

An embodied AI cell that sees, decides, and acts on the line

We built an embodied AI system that fuses multimodal perception, AI decision-making, and robotic motion control so a production cell can inspect parts, plan handling, and execute pick-and-place without manual reprogramming.

Embodied AI robot operating on a smart factory production line

+35%
Cell throughput per shift: -70%
Manual reprogramming time: 99.4%
Defect detection accuracy

Challenge

Each new SKU forced engineers to hand-tune vision models and robot trajectories for days.
Defect inspection relied on operators who could not keep up with line speed at peak hours.
Cell behavior was opaque: when something went wrong, no one could replay what the robot saw or decided.

Solution

Combined RGB and depth cameras with a vision-language model so the cell can recognize new parts from a few reference images.
Wrapped the existing robot controller in an agent that plans grasps, validates against safety policy, and falls back to a human approval queue on low confidence.
Added a run log that captures perception frames, decisions, and motion outcomes for every cycle, with replay for engineers.

Results

Cell throughput rose 35 percent on the pilot SKU family without changing hardware.
Onboarding a new SKU dropped from roughly three days of reprogramming to under four hours.
Defect detection accuracy reached 99.4 percent in production, with low-confidence cases escalated to a human reviewer.

Tech & integrations

ROS 2
Vision-language model
RGB-D cameras
PLC / robot controller bridge
MQTT telemetry
Run-log replay tooling

“For the first time the cell adapts to new parts on its own, and when it hesitates we can see exactly what it saw and why.”

Head of Manufacturing Engineering - Regional smart-factory operator

Discuss a similar engagement->