Inverse Constraint Learning: What It Is, Why It Is Different, and Why It Matters for Physical AI
Most AI safety systems tell machines what to do. Averiom's ICL approach learns what not to do from expert operator behaviour — a distinction that turns out to matter enormously in physical environments.
If you ask most people how AI safety works, they will describe something like this: you train a model on examples of correct behaviour, you deploy it, and it does the correct thing. When it does not, you add more examples or adjust the training.
This is a reasonable description of imitation learning — and it works well for a large class of problems. It does not work well for physical safety in industrial environments, and understanding why requires thinking carefully about what "correct behaviour" actually means when human skill is involved.
The Problem with Positive Examples
Consider teaching a robotic system to weld a structural joint. You collect data from expert welders. You train a model on their trajectories. You deploy.
The model learns to reproduce the average expert trajectory. On a standard joint, in standard material, under standard conditions, this works adequately.
Now the material has a slight surface inconsistency. Or the fixture is 0.3mm out of position. Or the ambient temperature is different and the puddle behaviour changes. The expert welder adapts — they have seen thousands of these variations and their response is embodied, automatic, largely unconscious.
The imitation model does not adapt. It reproduces the trained trajectory. The weld fails.
More importantly: there is no safety in this model. The model learned what correct looks like. It did not learn what unsafe looks like. It has no representation of the failure boundary — the region of action space that leads to bad outcomes. It cannot tell you whether a given trajectory is approaching danger. It can only tell you whether it looks like what the expert did.
What Inverse Constraint Learning Actually Does
ICL takes a different approach. Instead of asking "what did the expert do?", it asks: "what did the expert avoid, and why?"
The data source is not expert trajectories — it is correction events. The moments when an expert welder's hand makes a micro-correction: a slight course change, a speed adjustment, a pressure variation. These are not random. They are the physical expression of constraint knowledge — the expert's embodied model of where the failure boundary is.
ICL learns a probabilistic model of the Negative Manifold: the region of action space that expert operators consistently avoid. This is not a path. It is a boundary. The distinction matters enormously.
A path-based system can only be correct or incorrect — it reproduces the training trajectory or it deviates. A boundary-based system defines a safe operating envelope. Within the envelope, anything goes. The operator has full freedom of motion. At the envelope boundary, the system intervenes.
This is why Averiom's architecture is called "Physics > Probability" in our engineering philosophy. The constraints are not statistical guesses about what correct looks like. They are learned representations of the physical limits of safe operation — grounded in what real materials and real machinery actually tolerate.
The Technical Architecture
For those who want the mechanism: Averiom's ICL implementation operates in three phases.
Phase 1: Correction Event Detection. The sensor fusion layer (Vision, IMU, Force at 1kHz) identifies micro-corrections — defined as trajectory deviations that meet specific kinematic criteria: direction reversal or magnitude change above threshold within a short time window, correlated with continued task progression. These are not errors. They are the signal we want. They represent the operator's constraint knowledge expressing itself through action.
Phase 2: Negative Manifold Construction. Correction events are used to fit a boundary model in the relevant action space (tool position, velocity, force, or combinations thereof depending on the domain). We use a variant of Gaussian Process classification, chosen for its uncertainty quantification properties — the model does not just output a boundary; it outputs a boundary with confidence intervals. This matters for tiered intervention: the system can apply different response levels depending on how deep into the uncertain region the operator is moving.
Phase 3: Real-time Boundary Evaluation. The Prediction Kernel runs continuously at the edge. It evaluates the current trajectory against the learned Negative Manifold and uses a Kalman filter to forecast trajectory 150–300ms ahead. When the predicted trajectory intersects the manifold, the Inhibition Engine is triggered — haptic, damping, or hard lock depending on proximity and confidence.
The entire inference chain runs in under 5ms on the edge device. No cloud dependency. No network latency. The boundary model lives on the device; the constraint knowledge is local.
Why This Is Different from Rule-Based Safety
Traditional safety systems use rules: "if force exceeds X Newtons, stop." These rules are written by engineers based on material specifications and safety factors. They are conservative by design — they must account for worst-case conditions, operator error, measurement uncertainty, and legal liability.
The result is safety systems that trip frequently on legitimate operations. Every false positive is a stoppage, a frustration, a temptation to bypass. Safety systems that are too sensitive get defeated — physically, by removing sensors, or procedurally, by writing exceptions into the safe work procedure.
ICL-learned constraints are different because they are derived from what expert operators actually avoid in practice — not what the specification says they should avoid. The learned boundary accounts for material variability, tool wear, operator technique, and the actual tolerance of the physical process, because all of these are encoded in the correction events from which the model is learned.
This does not mean the constraints are less safe. The expert operators whose correction events train the model are, by definition, operating safely. What it means is that the constraints are calibrated to the actual failure boundary of the process — not to a conservative safety margin around an idealised model.
The Portable Profile
One aspect of ICL that has significant implications beyond individual deployments is the portability of the learned constraint model.
In current industrial practice, skilled worker knowledge is locked inside the individual. When a master welder retires, their constraint knowledge — developed over decades of physical experience — largely leaves with them. Junior workers must re-learn through their own correction events, their own near-misses, their own gradual accumulation of embodied skill.
Averiom's .avm profile is a portable representation of constraint knowledge. A profile trained on 50 expert welders in a specific composites fabrication domain captures the collective Negative Manifold of that domain. A new worker, on their first day with an unfamiliar material specification, operates inside an envelope derived from the combined experience of fifty experts.
This has immediate implications for workforce development — particularly relevant in the skilled trades shortfall Australia is currently experiencing. TAFE and industry training programmes face the challenge of transmitting embodied skill that has historically resisted formalisation. The .avm profile is not a replacement for hands-on training; it is a safety envelope that makes hands-on training safer and more efficient.
Where We Are and Where We Are Going
Averiom's ICL implementation is at Technology Readiness Level 4 — validated in laboratory conditions, with demonstrated end-to-end latency of 182.5ms in early hardware configurations and an architectural target of sub-5ms in the v1.0 production runtime.
We are exploring beachhead applications in two domains:
Autonomous mining equipment in Queensland, working through the Industry Growth Programme with advisors from the Department of Industry, Science and Resources, and in discussions with the Australian Manufacturing Capability Network (AMCN).
Composites welding and related trades — precision fabrication applications where the gap between current automation capability and expert human performance is most visible and most costly.
US Provisional Patent Application No. 63/962,640 covers the distributed edge architecture and constraint learning methodology.
We are seeking design partners who want to help validate this architecture in production environments, and investors who understand that the governance layer for physical AI is a foundational infrastructure bet — not a feature.
The safety system that actually stops the harm before it happens does not exist at scale yet. That is what we are building.
Talk to us about design partnerships or investment →
Related: Why 5ms is the line between a safety system and a logging system · The Cage Problem — 40 years of exclusion zones · Full peer-reviewed paper: Springer CCIS · HCII2026 · 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.