← All projects

Robotic Entropy Auditing

A Unitree Go2 robotics stack for converting factory-floor observations into entropy-based measures of organization, disorder, and run-to-run change.

Accepted · Journal of Intelligent Manufacturing.

Measuring factory-floor change with robot data.

This project uses a mobile robotic sensing stack to assess factory-floor organization from spatial, motion, and command-stream data.

The system converts costmaps, zone definitions, point clouds, odometry, and robot commands into normalized entropy metrics, anomaly scores, and run-level summaries for comparing factory-floor conditions over time.

Figures from the paper

What triggered an alert

Nominal and narrowed aisle photographs, with route-deviation and mean-JSD scores for three changed audits.
A narrowed aisle triggered a route-deviation alert, while one cabinet-removal run triggered a distributional alert. Scores are divided by each indicator’s calibration threshold; filled dots mark alerts. Aisle widths were reported by the operator.

Robot paths under aisle narrowing

Four maps comparing estimated robot paths under low and high narrowing of aisles A and B.
Gray paths show eight nominal evaluation audits; colored paths show three repetitions of each changed condition. Maps and paths are robot estimates, not surveyed positions or intervention boundaries.
Explore the system workflow

Acquisition and audit workflow

Workflow from saved map and robot measurements through offline entropy and route comparisons, calibration, and audit evidence.
Fresh observations are recorded on each audit and compared offline. Six reference runs, eight calibration runs, and 32 evaluation runs have separate roles. The dashed box shows proposed human follow-up.

What the audit measures

Robot sensing and navigation data

Uses costmaps, zones, point clouds, odometry, and command streams collected from a Unitree Go2 robotics workflow.

Entropy-based run summaries

Computes occupancy entropy, zone spatial entropy, point-cloud eigentropy, trajectory entropy, zone visit entropy, and command-stream entropy.

Run-to-run change detection

Produces Jensen-Shannon run comparisons, anomaly scores, and normalized AUC over time for comparing floor states and identifying disorder.