Autonomous ground inspection robots for industrial plants (legged & wheeled operator rounds)

Operators in chemical plants, cement works, refineries, substations and offshore platforms walk rounds to read gauges, listen for leaks and feel for hot bearings, often in hazardous or hard-to-reach places. Ground robots, legged ones such as ANYmal and Spot or wheeled platforms, now run these rounds autonomously on schedule, carry thermal, acoustic and gas sensors, read analogue instruments with vision models and push the results into maintenance systems. The benefit is more frequent, consistent data and fewer people in hazardous zones.

Problem & value

• Today: shift operators walk fixed rounds on paper or tablet checklists. Coverage is limited by shift patterns, access (stairs, grated floors, confined areas) and hazards (explosive atmospheres, heat, noise, high voltage). Many gauges are analogue and not worth retrofitting with sensors.
• AI value: the robot repeats the same route with the same sensor poses, so readings are comparable over months. Vision models read analogue gauges and sight glasses; thermal and acoustic baselines expose deviations earlier than a human round; results land as structured data instead of paper notes. Staff are removed from the most hazardous locations, not from the plant.
• Evidence from deployments (vendor-reported): ANYbotics reports more than 33,000 inspections by ANYmal at the Vigier cement plant and about 200 units in service at industrial customers; BASF has evaluated ATEX Zone 1 robots in chemical-plant conditions. Treat these as vendor figures and ask references for false-alarm rates and uptime before sizing a business case [1][2].

Variants

• Chemical, petrochemical, oil and gas (explosive atmospheres): ATEX/IECEx-certified robots for Zone 1/2 (gas) and Zone 21/22 (dust) areas; gas and acoustic leak detection. This covers the merged variant "autonomous robotic inspection rounds in industrial plants", including the BASF evaluation and ATEX Zone 1 evidence.
• Cement, steel, mining: dust and heat; thermal checks of kilns, conveyors, bearings.
• Power and utilities: substations and switchgear, with thermal imaging and acoustic partial-discharge detection; the ANYmal page cites RTE substation inspection [1].
• Offshore platforms: reduced-crew operation with remote supervision.

How it works

1. mission plan (waypoints, sensor actions per asset, schedule)
2. autonomous navigation (LiDAR/visual SLAM, obstacle avoidance, stairs, doors)
3. capture at each waypoint (RGB/zoom, thermal, ultrasonic microphone, gas sensors)
4. on-robot/edge analysis (gauge reading, thermal ROI temperature, acoustic anomaly score)
5. fleet software (e.g. Orbit, Yokogawa OpreX, ANYbotics software)
6. baselines and trends
7. alert/work-order proposal in CMMS/APM
8. human review and decision
9. docking, charging, next mission

• Navigation: the robot builds a map on a first teleoperated run, then replays the mission. No GPS needed indoors. Legged robots handle open grated stairs, low crawl spaces and uneven or wet floors [1].
• Perception: a fixed-pose capture makes analysis simpler. Gauge reading uses a detector to find the dial, keypoint or segmentation of the needle and a calibration to the scale; VLMs add reading of sight glasses, tags and digital displays. Thermal regions and acoustic spectra are compared with asset baselines; unsupervised anomaly detection flags what thresholds miss.
• Sensors (ANYmal example): 20x optical zoom camera, thermal camera (-40 to 550 degrees C), pan-tilt unit, ultrasonic microphone (0-384 kHz), 360 degree LiDAR, optional acoustic gas-leak imager and gas concentration sensors, about 90 minutes of runtime and automatic docking [1].
• Human-in-the-loop: operators supervise missions, approve work orders and teleoperate when stuck; shutdown decisions never rest on AI readings alone.

Data preparation

The robot needs a plant map, inspection routes and points, reference images and readings for each asset, and links to maintenance and asset data. These live in CMMS/EAM or ERP (asset master), P&ID drawings and PDFs on file shares, SCADA/historian systems, inspection round sheets on paper and in the knowledge of experienced operators.

Typical gaps: missing or inconsistent asset IDs between CMMS, SCADA and drawings, outdated plans, paper-based round logs, no historic images of normal and faulty states, no defined thresholds for gauges and thermal readings, and unclear access rights between OT and IT networks.

Preparation steps: harmonise asset IDs and master data (PREP-005, PREP-011), digitise round sheets and drawings (PREP-006, PREP-008), define inspection points and finding categories with the maintenance team (PREP-019), record baseline images and readings over several rounds and label deviations with experts (PREP-018, PREP-020), and set up read-only data exports from historian and CMMS. Clarify OT/IT network zones early with IT security.

Rough share of effort: about 30 percent, mostly asset data and baseline collection. See GUIDE-003; for sensor time series also DATA-003.

Implementation paths

Which path for whom

• Enterprise: Path A, because established inspection-robot platforms with certified hardware and ready inspection software exist, which fits commodity rounds in plants; Path B or C only for strategic inspection analytics.
• Developer: Path C, because ROS 2, Nav2 and open vision models run on HW1 edge hardware and allow building and adapting own inspection software.

Path A — Buy (recommended)

• Legged: ANYbotics ANYmal (Zurich, Switzerland, EU-adjacent; ATEX/IECEx variants for hazardous areas; Yokogawa has announced integration of its OpreX robot management software [3]), Boston Dynamics Spot with Orbit and the Spot SDK (US; industrial inspection with thermal and visual payloads [2]), Unitree and Deep Robotics (China; cheaper, check cybersecurity and supply-chain policy for critical infrastructure).
• Wheeled/software (EU): Taurob (Austria, ATEX-oriented), Energy Robotics (Germany, multi-brand software layer integrating SAP PM, Maximo, AVEVA PI).
• Choose A when: you want a supported, certified system. Explosion protection certificates, safety cases and service contracts matter more than model novelty.

Path B — Build on cloud APIs

• Use the robot vendor's SDK (e.g. Spot SDK) for missions and data capture, then send images to a multimodal model (Gemini, GPT, Claude, Mistral) with a JSON schema for gauge, valve and tag reading. Use EU regions and send only cropped asset images, without people.
• Good for pilots and odd instruments; latency and connectivity make it unsuitable for safety functions.

Path C — Self-host (open source)

• Robot middleware: ROS 2 with Nav2 for navigation, slam_toolbox or RTAB-Map for mapping, OpenVINS or similar for visual-inertial odometry. NVIDIA Isaac ROS for GPU-accelerated perception.
• Perception: YOLO-class detectors, SAM-family segmentation, open VLMs (Qwen-VL family) served on-site; open-source anomaly-detection libraries such as anomalib for thermal/visual baselines.
• Hardware: HW1 on the robot (Jetson Orin class); HW2 server on site for VLM and reporting.
• Effort: high; you own safety and certification, so only with a robotics team.

Path D — Custom training

Fine-tune gauge readers on your own dial types (hundreds of robot-collected images each) and acoustic/thermal classifiers per asset class. Rare faults need simulated data (ROBO-010).

State of the art (2026)

• Production readiness: legged robots run routinely at selected plants; most sites still pilot 1-5 robots.
• Autonomy: repeatable missions and docking are standard; elevators, narrow pipe racks and heavy rain remain hard.
• Perception: VLMs improved gauge and label reading and cut per-gauge calibration effort, but glare, dirt, odd scales and parallax still cause misreads. Use a confidence threshold and a second capture before alerting.
• Limits: battery runtime of about 90 minutes needs docking stations along long routes; radio coverage in steel structures; robots can block escape routes or trigger alarms; ATEX certification limits payloads and sensors; false positives erode operator trust fast.

Shared resources & combinations

• Same perception stack (object-detection, image-classification, vlm, anomaly-detection): VIS-001, VIS-003 (drone inspection), VIS-043 (pipe CCTV), VIS-007 (PPE monitoring).
• Data consumers: PRED-006 predictive maintenance uses the robot as a data source; ANOM-013 leak detection; maintenance work-order flow in CMMS.
• Robotics stack: ROBO-002 AMRs (shared navigation and fleet management), ROBO-006 drones (combine aerial and ground for tall structures), ROBO-010 simulation for training and mission testing, ROBO-016 safety functions for robots working near people, ROBO-004 for future manipulation.

Effort, cost & KPIs

• Path A: pilot in 2-4 months (route mapping, integration, safety assessment); robots are a six-figure investment per unit with subscription, so contact vendors for quotes; ATEX variants cost more.
• Path B/C: 6-12+ months with a robotics team.
• Cost drivers: robot, dock, payloads, CMMS/APM integration, safety case, network coverage.
• KPIs: mission completion rate, availability, inspections per week, anomalies found vs false alarms, gauge-reading accuracy against manual, time from detection to work order, unplanned downtime avoided, person-hours in hazardous zones.

People & change

Involve management, the plant or maintenance lead (process owner), operators and inspectors who do rounds today, IT and especially IT security for the OT network, the data owners of CMMS and SCADA data, and the works council from day one (GUIDE-001).

Works council: yes. Autonomous robots with cameras move through workplaces and record people and activities; logs of routes, findings and responses can be suited to monitor behaviour (§87(1) no. 6 BetrVG, objective suitability suffices). Inform the works council in good time about AI-planned work processes (§90(1) no. 3), consider an expert (§80(3)), and conclude a works agreement covering camera zones, retention and prohibition of performance evaluation (GUIDE-004). Public operators involve the Personalrat.

Data protection: persons in images are personal data; involve the DPO, use blurring of faces or exclusion zones, define purpose and retention and run a DPIA where recording is systematic (GUIDE-005).

AI literacy: support operators with training on reading robot findings, false alarms and manual override (GUIDE-006). Typical concerns are surveillance and job loss; explain that the robot covers hazardous or monotonous rounds, that inspectors move toward analysis and decisions, and that findings are suggestions, not verdicts (GUIDE-008). Pilot on one area with volunteers, measure detection quality and false alarms (GUIDE-009), and keep humans in control by letting staff confirm findings, trigger work orders and stop the robot at any time.

Risks, regulation & compliance

• Explosion protection: use in hazardous areas requires equipment certified under Directive 2014/34/EU (ATEX) and, internationally, IECEx; the operator's explosion-protection document under Directive 1999/92/EC and zone classification must cover the robot, its charging dock and payloads [4].
• Machinery: the robot is machinery; the Machinery Regulation (EU) 2023/1230 applies from 20 January 2027 and replaces the Machinery Directive, with specific rules for safety components with self-evolving behaviour. Plan CE conformity and risk assessment accordingly [5].
• EU AI Act: classification depends on use. Pure condition monitoring of an industrial plant is usually minimal risk. AI acting as a safety component of machinery (Annex I) or of critical infrastructure (Annex III) can be high-risk; check each case. No Art. 5 issue arises if the robot is not used for emotion recognition of workers or for biometric categorisation; do not repurpose the robot as a surveillance tool for staff [6].
• GDPR and works councils: cameras and microphones will capture workers. Define no-go zones, blur people, keep retention short, run a DPIA and involve the works council (DE: section 87 BetrVG).
• Cybersecurity: apply IEC 62443 segmentation and NIS2 duties; review the supply chain of non-EU robots.
• Operational risks: collisions, blocked escape routes, misread gauges. Use geofences, e-stop, fixed sensors for safety-critical points; the robot supplements safety systems.

Getting started

1. List the rounds: assets, instruments, frequency, hazards. Pick one route with 20-50 checkpoints.
2. Check access (stairs, doors, floor type, zones) and radio coverage; confirm the zone classification.
3. Run a 3-month pilot with one vendor robot, comparing robot readings against operators' manual readings.
4. Integrate alerts with CMMS/APM, with human approval for work orders; agree on retention and DPIA.
5. Measure KPIs and decide on roll-out, additional docks and further sensors.

Sources

1. ANYbotics — ANYmal inspection robot (sensors, docking, industrial use) — https://www.anybotics.com/robotics/anymal/
2. Boston Dynamics — Spot product page (Orbit, SDK, industrial inspection) — https://bostondynamics.com/products/spot/
3. ANYbotics — Yokogawa OpreX integration with ANYmal — https://www.anybotics.com/news/yokogawa-and-anybotics-to-integrate-oprex-robot-management-core-software-with-anymal-robotic-inspection-solutions/
4. Directive 2014/34/EU (ATEX equipment) — https://eur-lex.europa.eu/eli/dir/2014/34/oj
5. Regulation (EU) 2023/1230 (Machinery Regulation) — https://eur-lex.europa.eu/eli/reg/2023/1230/oj
6. Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj
7. Nav2 (ROS 2 navigation) — https://docs.nav2.org
8. anomalib (open-source anomaly detection) — https://github.com/openvinotoolkit/anomalib