The Cage Problem: Why Industrial Safety Has Been Trading Productivity for the Illusion of Control
For 40 years, the dominant paradigm in industrial safety has been exclusion zones and hard stops. This has kept workers safe by keeping them away from the work. It is time for something better.
In 1979, a Ford Motor Company worker named Robert Williams became the first human killed by an industrial robot. The response from the industry was, predictably, a cage. Put a fence around the robot. Keep the humans out. Problem solved.
For the next four decades, this paradigm — exclusion through separation — became the bedrock of industrial safety. ISO 10218, the primary standard governing industrial robot safety, is fundamentally a document about fences: how far from the work envelope, what kind of barrier, how fast the hard stop must engage when the boundary is breached.
This approach has saved lives. It has also, quietly, become the ceiling on what industrial automation can achieve.
What the Cage Actually Does
A safety cage does two things simultaneously. It protects the human from the machine. And it prevents the human from being near the work.
For many applications, this is fine. For others, it is a fundamental constraint on quality, productivity, and the nature of the work itself.
Welding is the clearest example. The best welds — particularly in composites, structural steel, and precision fabrication — require human presence and human judgment. The welder reads the puddle. They adjust travel speed by feel. They compensate for material inconsistency in real time. This is tacit knowledge — knowledge that lives in the hands, not in a manual. Caging this work means automating it, which means losing access to that knowledge. Current robotic welding systems produce consistent, adequate welds. They do not produce expert welds on variable or non-uniform material.
Mining equipment operation presents a different variant of the same problem. The transition to autonomous haul trucks has improved safety metrics in open-cut operations. It has also removed the operator's embodied knowledge of the machine — the subtle vibration patterns that signal a hydraulic anomaly before any sensor trips, the learned feel of load distribution on variable terrain. That knowledge is not gone; it lives in the operators. But the current architecture of autonomous systems has no mechanism to encode or use it.
The False Choice
The industry has implicitly accepted a false binary: either you automate (and lose human skill), or you keep humans in the loop (and accept risk).
Averiom exists to reject this binary.
The insight at the core of our architecture is that human skill and machine safety are not opposites. They are complementary. The expert welder is not dangerous because they are human — they are skilled because they are human. The source of their skill (embodied, adaptive, context-sensitive judgment) is also the source of the risk (humans make errors, especially under fatigue, distraction, or unfamiliar conditions).
What you actually want is a system that captures the constraint knowledge of the expert and enforces it on everyone — including the expert, on their worst day.
This is what Inverse Constraint Learning does. Not: "here is the correct path." Instead: "here are the regions where errors happen — the Negative Manifold — learned from watching experts correct themselves."
The Productivity Cost Nobody Measures
Every safety cage, every exclusion zone, every hard-stop interlock carries a productivity cost that is rarely measured in safety ROI calculations.
The most visible cost is downtime: when a safety system trips, work stops. In a cage-based paradigm, false positives are common because the system has no model of intent — it only knows that a boundary was crossed, not why. An operator reaching into the work area to adjust a fixture trips the same stop as an operator about to be struck. The system cannot distinguish.
The less visible cost is workflow design. Safe work procedures around exclusion zones require elaborate choreography: safe work positions, lockout/tagout protocols, re-entry sequences. These procedures are necessary. They are also time-consuming, and in high-frequency-access work environments, they become a source of their own risk as workers find shortcuts.
The deepest cost is capability limitation. There are classes of work that simply cannot be done safely under current paradigms without removing the human entirely. The composites welding applications we are exploring with AMCN partners involve precision work in confined geometries where both full automation and human-only approaches have clear failure modes. The gap between them is exactly where Averiom operates.
What Governed Agency Looks Like
The alternative to the cage is not the absence of control. It is better control — control that is proportional, continuous, and invisible unless needed.
Imagine a composites welder working on a structural frame. Their tool's force and trajectory are being monitored at 1kHz. The Negative Manifold for this specific joint geometry — learned from 10,000 correction events across 50 expert welders — is loaded into the edge device. The welder works normally. They do not feel the system unless they approach a constraint boundary.
When they do, the response is graduated. First, a haptic signal through the tool handle — a slight vibration, like a rumble strip. Most experienced operators will self-correct here. The system has communicated without interrupting. If the trajectory continues toward the violation, resistance increases — viscous damping, like working through slightly thicker medium. The tool still moves; the operator is not stopped. They are guided.
Only in the case of a genuine imminent violation — predicted to breach the hard limit within the next 150ms — does the system engage a hard lock.
And every one of these events is signed, timestamped, and stored in the operator's .avm profile. Not just for compliance. For training, for improvement, for the portable reputation that should travel with a skilled worker across their entire career.
The Mining Application
In autonomous and semi-autonomous mining equipment, the cage problem manifests differently. The operator is already physically separated from the hazard — they sit in a cab or operate remotely. But the constraint knowledge problem is the same.
Current autonomous systems operate on programmed paths and sensor-triggered stops. They are brittle on variable terrain, in conditions that differ from their training distribution, and in situations that require the kind of adaptive judgment that experienced operators carry.
The opportunity is not to replace the operator — it is to encode what the best operators know and enforce it as a real-time constraint on autonomous and semi-autonomous systems. The Averiom architecture is hardware-agnostic by design: the same Prediction Kernel and Inhibition Engine that governs a welding torch can govern a hydraulic arm, with domain-specific constraint models loaded from the relevant operator corpus.
We are seeking design partnerships with mining equipment operators and OEMs in Queensland to validate this architecture in the field. If your operation has both skilled operators and an automation programme, you have exactly the data we need.
What Comes Next
The cage had a 40-year run. It was the right answer to the wrong version of the problem.
The right version of the problem is not "how do we keep humans away from dangerous machines?" It is "how do we let skilled humans work alongside powerful machines, with the safety of both guaranteed by physics rather than distance?"
That is what we are building. Design partners and funding discussions welcome — get in touch.
Related: Why 5ms is the line between a safety system and a logging system · ICL explained — the technical mechanism behind the Negative Manifold · Watch the POC demonstration
We are seeking design partners in autonomous mining equipment and composites welding, and investors who understand the foundational infrastructure opportunity in physical AI governance. Supported by the Australian Government Industry Growth Programme and in discussions with AMCN and AMGC.