YardVision
Vision-only crane and yard intelligence — on a deliberate trajectory from passive monitoring to remote yard operation. Monitoring today; the physical no-man zone and gate interlock are the path to the rest.
The Thesis
Safety is physical. That is exactly why vision-only works.
Remote and autonomous yard operation were historically blocked by asking the perception sensor — radar, laser, vision — to also be the safety sensor: to prove no human hazard in an occupied yard, in dust and glare. That is an unbounded problem, and where radar and laser get costly, fooled, and still uncertain.
The old framing
Ask the camera to prove a negative — “there is no person anywhere in this occupied yard” — under dust, glare, and occlusion. Unbounded, expensive, and never quite certain. Every remote-operation attempt stalls here.
YardVision inverts it
The safety function is mechanical — a sealed no-man cell with gate interlocks. The hard “prove no human present” job goes to a gate switch. Vision is freed to do only what it is excellent at: machine geometry, load position, and clearances between equipment. No people in the frame to get wrong.
Vision-only is an architecture choice, not a limitation.
The Trajectory
Five rungs. Each a deliverable. Each gates the next.
Rungs 1–2 are present capability. Rungs 3–5 are destinations the architecture is built toward — attached to a real product, never claimed as present-tense.
Passive monitoring
Today · in design → deployFixed-optical crane position and load monitoring across the slab and coil yard. Sense, log, alert. Open-loop, human in command.
Operator assist
Near-termClearance and collision warnings and position guidance to the cab. The human is still in command — the system advises, it does not act.
Physical no-man zone + gate interlock
Prerequisite · roadmapBarriers and a defined sealed cell; locked, monitored, interlocked gates. Gate-open → all-crane-stop — hardwired, fail-safe, perception-independent. Open the gate and the crane is dead whether or not anything is detected inside. This rung gates everything above it.
Remote operation
Architected destinationCabin-out control of a verified-empty, sealed cell — unlocked only once the interlock holds. Vision is the operator’s eyes for machine and load tasks; the gate interlock owns human safety.
Supervised / bounded autonomy
Architected destinationThe system executes routine repetitive moves inside the manless interlocked cell; a human supervises and can override. Bounded, low-consequence yard moves first.
The Architecture
Two layers, deliberately separated
The life-safety layer is a mechanism. The reasoning layer is freed from carrying safety. Keeping them apart is the credibility of the whole design.
Layer 1
Safety interlock — boring on purpose, highest integrity
- Gate-state — not person-detection — is the interlock. Gate open → stop, regardless of whether anyone is seen. The safety case rests on a door switch proving a positive (“boundary closed”), not on perception proving a negative (“no person present”).
- Fail-safe by construction: sensor loss or ambiguity → assume unsafe → stop. A false positive costs throughput; only a false negative is dangerous — so the system is biased to over-stop.
- Independent of the perception and reasoning stack: if vision has a bad day, the interlock still fires.
- Person-detection inside the cell, where present, is an additive second layer — never the thing life depends on.
Layer 2
Vision operation — unburdened
- YardVision fixed-optical perception runs crane geometry, load position, hook alignment, and equipment clearances inside the verified-empty cell.
- Tractable precisely because the cell contains no humans — no people in the frame to get wrong.
- Recognition premise, fixed parallel eyes, no PTZ — the same thesis as HookVision.
The Governing Metric
Zone integrity, not detection accuracy
The metric that governs safety is the reliability of the boundary and the gate interlock — not how well a camera spots a person. Vision performance is stated for the operation task (machine and load), explicitly not for human-safety.
Boundary completeness
Is the cell fully sealed — every edge of the no-man zone accounted for?
Gate-state monitoring
Continuous, positive confirmation of every interlocked gate’s closed state.
Detection-to-halt latency
Gate-open to all-crane-stop, hardwired. {{TODO: confirm target}}
Fail-safe coverage
Every loss-of-signal and ambiguity path resolves to stop, not run.
Operational Honesty
The empty state has to be achieved — and re-achieved
The manless rule is absolute and never softened. But a working steel yard must reach and re-reach the empty state — maintenance, breakdowns, sling changes, the rigger who occasionally must enter.
Maintenance / lockout-access mode
Gate open means the crane is locked out — not merely stopped — with explicit re-verification required before resumption. This is not an exception to the safety rule; it is the operational design that makes the absolute rule hold in a real plant.
Why This Wins
Physical interlock for safety. Vision for operation.
Against the usual “vision + radar + laser for redundant safety” stack, the inversion is the argument that makes a safety head sign.
Fewer sensors
No radar or laser safety array to buy, calibrate, and maintain against dust and glare.
Lower cost
The life-safety layer is a gate switch and a barrier, not a perception system carrying the whole risk.
Higher integrity
Safety rests on a mechanism plant teams already trust — gate interlocks — not a perception claim they must take on faith.
The same operator-in-cabin reasoning runs through our crane-fatigue work — geometry, cognition, and the load, instrumented together.