The Problem Is AI Moving Into The Site
Caterpillar announced on September 2, 2026 that it is collaborating with FieldAI on physical AI, autonomy, and robotics for jobsites and factories. The September 6 newsflash treated that as part of AI moving from screens into operational environments.
That shift changes the acceptance question. A software feature can be rolled back quickly, but a robot, autonomous inspection system, or equipment-assist workflow has to perform around people, vehicles, weather, dust, noise, blocked routes, and changing work plans.
Why Jobsites Break Lab Assumptions
Jobsites change by the hour. A route that was clear in the morning may have a parked truck, trench, material stack, temporary barrier, or worker crew by afternoon.
Physical AI systems also depend on sensors, maps, radio coverage, battery state, maintenance condition, and human handoff. A digital twin can improve planning, but it is still a representation that must be checked against the live site.
The Cost Of Skipping Acceptance
Skipping acceptance can create downtime, damaged equipment, unsafe interactions, worker distrust, and expensive manual rescue. The first failure may not be dramatic; it may be a system that stops too often, misses context, or needs constant operator attention.
A practical cost estimate is robot-hours planned times stoppage rate times labor and equipment hourly cost. The formula should be paired with safety review, because a cheap stoppage is still unacceptable if it occurs in the wrong interaction zone.
How To Diagnose Site Readiness
Map routes, exclusion zones, shared travel paths, blind spots, radio dead zones, worker crossings, emergency-stop access, maintenance areas, and manual takeover points. The readiness review should include operators and frontline supervisors, not only vendor engineers.
Then list the assumptions the AI system makes about the site. If the system assumes a clear route, stable map, readable signs, reliable connectivity, or predictable human movement, each assumption needs a test.
Options For Deployment Levels
The lowest level is observation only, such as inspection images or site mapping with no autonomous interaction. The next levels are remote assist, supervised autonomy, task-specific autonomy, and full workflow integration.
Each level needs its own acceptance threshold. A camera inspection route can tolerate different failures than a machine operating near moving vehicles, suspended loads, or workers on foot.
Build The Site Acceptance Test
The test should include a pre-start survey, digital-twin check, marked route trial, worker-detection test, vehicle-interaction test, emergency-stop test, connectivity check, manual takeover drill, maintenance inspection, incident logging, and pass-fail threshold. Record evidence for each step.
The test should also define what happens after a site change. New routes, new equipment, major weather conditions, sensor replacement, software updates, and work-zone changes should trigger partial retesting before the system resumes normal operations.
A Worked Example
A quarry wants an autonomous inspection robot to monitor conveyors and stockpile areas. The first acceptance route avoids active haul roads, logs communication gaps, tests worker detection at crossings, and requires manual recovery within a defined time.
The system fails acceptance if it loses location near vehicle traffic, cannot stop reliably around a worker crossing, or leaves operators unclear about takeover responsibility. Passing the route does not authorize unrelated routes; each operating envelope gets its own record.
Measures That Prove Control
Useful measures include interventions per operating hour, mission completion rate, false stops, near misses, communication gaps, manual recovery time, battery and maintenance exceptions, worker reports, and retests after site changes. These measures should be reviewed with safety and operations together.
Also measure whether incidents lead to updated tests. If the acceptance checklist never changes after field problems, the organization is certifying yesterday's assumptions instead of today's site.
The Next Step This Week
Choose one bounded route or task and write the acceptance test before expanding the system. The test should name the operating envelope, pass-fail threshold, stop authority, evidence owner, and retest triggers.
SynHy would treat this as a live operating artifact. It belongs in the same review path as safety plans, shift handoff, maintenance records, and site-change notices.
Sources And Methodology
This article was triggered by Caterpillar's announcement that it is working with FieldAI to advance AI-powered industrial innovation. It also uses FieldAI's discussion of industrial AI and NVIDIA Omniverse digital twins.
The acceptance test is SynHy original analysis informed by OSHA's construction struck-by hazard guidance, ISO's ISO 3691-4:2023 driverless industrial truck standard page, and NIOSH's struck-by injury prevention bulletin. It is not a substitute for site-specific engineering, legal, or safety review.