Contact usResearch
The intellectual foundation of Averiom’s Inverse Constraint Learning technology, validated through peer-reviewed research and published in indexed academic proceedings.
Inverse Constraint Learning for Real-Time Manufacturing Safety: A Predictive Intervention Framework
Manufacturing defect rates of 3–8% in advanced composites operations result in costly rework and constrain critical aerospace and defence production capacity. Current quality assurance approaches rely on post-production inspection or operator training, neither of which prevents errors during execution.
This work presents a novel real-time predictive safety system that prevents manufacturing defects before they occur using inverse constraint learning. Unlike traditional supervised learning approaches that teach “correct paths,” our system learns the “negative manifold” — boundaries of forbidden outcomes — from expert operator behaviour. By defining what must not happen rather than prescribing exact procedures, the system preserves human adaptability while enforcing deterministic safety boundaries.
The system architecture fuses multi-modal sensor data (computer vision, inertial measurement units, force sensors) to predict near-future trajectory states within a 150-millisecond horizon. Predicted states are evaluated against learned constraint manifolds in real time, triggering graduated physical intervention (haptic feedback → resistive damping → hard interlock) when boundary violations are detected.
Proof-of-concept validation on controlled manufacturing tasks demonstrates: 182.5ms end-to-end latency (perception → prediction → intervention), zero false negatives across 15 violation events (100% detection rate), 89.6% tracking uptime with autonomous recovery from occlusion, deterministic hardware intervention capability via servo actuation, and complete cryptographically-signable audit trail generation.
The inverse constraint learning paradigm addresses three critical challenges in human-machine safety: (1) Adaptability — preserves operator dexterity for variable tasks while enforcing non-negotiable boundaries; (2) Transferability — expert knowledge encoded as versioned constraint profiles enables skill democratisation; (3) Verifiability — negative space representations enable formal safety verification for certification.
Initial deployment targets include aerospace composites welding and carbon fibre layup operations for AUKUS submarine manufacturing. The framework generalises to surgical guidance systems, high-voltage maintenance, collaborative robotics, and any domain requiring real-time physical safety governance in human-operated or semi-autonomous systems.
- Ng, A. Y., & Russell, S. J. (2000). Algorithms for inverse reinforcement learning. Proceedings of the 17th International Conference on Machine Learning, 663–670.
- Haddadin, S., Johannsmeier, L., & Ledezma, F. D. (2019). Tactile robots as a central embodiment of the Tactile Internet. Proceedings of the IEEE, 107(2), 471–487.
- Lasota, P. A., Fong, T., & Shah, J. A. (2017). A survey of methods for safe human-robot interaction. Foundations and Trends in Robotics, 5(3), 261–349.
- ISO 13849-1:2023. Safety of machinery — Safety-related parts of control systems — Part 1: General principles for design. International Organization for Standardization.
NOTE: Reference [4] in the submitted abstract cites the Averiom POC repository. Cite as: Mishra, T. (2026). Averiom Predictive Safety System: Proof-of-Concept Validation. Technical Report. Intelligrate Pty Ltd, Brisbane, AU.
“This is a strong and ambitious poster proposal with a compelling central idea and clear industrial relevance. Its primary contribution lies in reframing safety learning around negative space and embedding that insight into a tightly engineered real-time intervention system. While the empirical validation is necessarily preliminary and the theoretical framing could be deepened, the work is well suited for a poster venue and has clear potential to mature into a high-impact systems or interdisciplinary journal contribution.”
DOI will be assigned upon final publication. Paper ID: 6491.
Future directions
Full paper submission to a systems or robotics journal expanding the theoretical framework and deepening empirical validation with field deployment data.
Design-partner pilot studies in composites welding and autonomous mining to generate production-grade validation data for peer review.
Conversion of US Provisional Patent Application No. 63/962,640 to non-provisional (deadline Jan 2027), with Australian provisional filing in parallel.
Interested in the research or exploring deployment?
We are seeking design partners to generate field validation data in composites welding and autonomous mining, and investors to fund the transition from TRL 4 to production deployment.