Autonomous patrol and thermal inspection with a quadruped robot at an electricity distribution site
Running on site in limited scope and under supervision. Scaling depends on the pilot outcome.
- Industry
- Electricity distribution and energy infrastructure
- Customer
- Electricity distribution company, Turkey
- Period
- 2026-06 → sürüyor
- Operating mode
- Supervised autonomy
Environment
Where the robot works
A facility building with the charging dock plus the open yard around it. Concrete and asphalt ground, ramps and stairs, variable night lighting, with staff and vehicle traffic on site.
Problem
Operational problem
Periodic thermal checks and night patrols depended on human shifts: the inspection interval slipped with shift load, and two measurements of the same point were taken by different people from different angles, so they were not comparable. Records were mostly handwritten notes and phone photos.
Task
What the robot does
- 1Leaves the dock, drives the taught route autonomously and stops at the inspection points along it.
- 2At each point it turns the gimbal camera to the taught angle and zoom, then captures the optical and radiometric thermal image from the same framing.
- 3For two-state targets such as lights, doors and fire doors it compares the image with the taught references and classifies the state.
- 4When staff are seen it reports helmet status in three values: helmet on, no helmet, undetermined.
- 5Findings are written to the dashboard with a timestamp, position and image; the robot returns to the dock when the mission ends.
Hardware
Hardware in use
Unitree Go2-W EDU
Wheeled-legged quadruped platform; ramp and stair traversal verified in the stock gait mode.
Livox Mid-360
3D LiDAR: the primary source for mapping and localization.
Intel RealSense D435i / D455
Near and mid-range depth; the blind zone at the robot's feet is covered together with the LiDAR.
InfiRay T3 Pro
Radiometric thermal camera; read in raw 16-bit mode with an FFC applied at start-up.
SIYI A8
Pan-tilt-zoom gimbal camera with an RTSP H.265 stream; aimed at the target at each inspection point.
Robot üzeri işlem birimi
Navigation and perception run onboard the robot; the dashboard runs on a separate computer.
Şarj istasyonu
Start and end point of every mission; localization is anchored to this dock.
Components used in the installation.
Software
Software and AI
FAST-LIO
LiDAR-inertial odometry; the map and the live pose come from it.
Şarj istasyonuna göre konumlanma
Instead of a drifting map-localization chain, a frame anchored to the dock is used; the lock solution must pass a vertical-consistency gate.
Nav2
Path planning and mission execution.
YOLO11 + TensorRT
FP16 inference; person detection and tracking, running together with the helmet model.
Öğretilmiş denetim noktası
The operator boxes the target and teaches both states; on patrol the image is compared against these references.
İSTRİCH panosu
Sensor streams, mission control, the finding log and the audit trail in one interface; the services survive a reboot.
Workflow
Mission flow
- 01
Dock
The robot waits, the system is disarmed.
- 02
Mission start
The operator arms it, localization locks to the dock.
- 03
Autonomous route
Inspection points are visited in order.
- 04
Point inspection
The gimbal turns to the target; optical and thermal frames are captured.
- 05
Classification
State comparison, thermal threshold, helmet status.
- 06
Finding and alert
The record goes to the dashboard, a breach notifies the operator.
- 07
Return to dock
The mission ends and the system is disarmed again.
The loop is closed: every mission starts and ends at the dock. The operator can stop it at any step.
- Dock: the robot waits, the system is DISARMED.
- The operator starts the mission, the system is ARMED and localization locks to the dock.
- Autonomous route: the robot visits the inspection points in order.
- It stops at the point, the gimbal turns to the target, optical and thermal frames are captured.
- Classification: state comparison, thermal threshold check, helmet status.
- The finding is written to the dashboard; a threshold breach notifies the operator.
- The mission ends, the robot returns to the dock and the system is disarmed again.
Human
The human role
Supervised autonomy
The operator starts the mission and watches it from the dashboard; taking over driving, emergency stop and mission abort stay with the operator at all times. The robot never starts a mission by itself and never closes a finding by itself: every finding goes to the operator.
Safety
Safety approach
- Disarmed by default: the system accepts no motion command at start-up.
- Emergency stop and speed limit from the dashboard; the robot stops if the drive command is lost.
- Virtual fence: a route leaving the defined area is rejected.
- Geometry gate: if the robot's swept footprint does not fit the passage, the mission does not start.
- Standing up and lying down happen only in the verified mode chain; commands from intermediate modes are refused.
Limits
What the system cannot do
- The wheeled variant has no dynamic obstacle avoidance: an unexpected obstacle on the route needs operator intervention.
- Localization depends on map coverage. On a thin map the lock score drops, so the measured coverage value was turned into a gate and a mission that fails it does not start.
- Multi-floor autonomy is not solved: each floor needs its own map and a re-lock on arrival, and elevator integration has not been done.
- Face recognition did not hold up on site and was switched off. On real staff photos the median similarity for the same person measured 0.31, which is not a reliable acceptance level. Instead of identity, only helmet and state inspection are used.
- The thermal camera needs a flat-field correction at start-up; skipping it makes readings systematically low. We report differences and trends rather than absolute temperature claims.
- The helmet model could not be tested enough at close range: no no-helmet reference footage was collected on site, so the false-negative rate in that condition is unknown.
- The ROS version on the robot is locked to an older branch; the move to the current branch is planned but not done.
Results
Measured results
- Driving, mission start and finding logging were run end to end from the dashboard; the system was installed as services that survive a reboot.
- Light, door and fire-door inspection ran live from taught points. Since an off-the-shelf detection model does not contain these classes, the method was switched to reference comparison.
- Map quality was made measurable: a wall-thickness measure is compared against the vendor localization threshold, and the new map covers twice the area of the previous one.
- Dashboard latency was reduced: endpoint responses moved from the 15-25 ms band to the 0.4-0.8 ms band (on-demand production, connection reuse and TCP_NODELAY).
Measurements come from commissioning and field tests; customer operational data is out of scope.
Platform
The robot used
Values are manufacturer data.
Site survey for a similar deployment
This deployment was configured around a single site's conditions. Floor, route, network coverage, charging point and safety rules differ from site to site; the scope is set by the survey report.