Embedded Edge-Intelligence for Contact-State Estimation

Project summary

Reading contact you cannot see.

Constrained nuclear maintenance tasks often encounter failure at the point where contact must be interpreted before it can be visually confirmed. Blind insertion, seal engagement, contact-guided retrieval and confined manipulation are typical examples, where the controller must judge whether contact is stable and whether seating is complete within restricted geometries and under limited visibility.

When this judgement is poor, minor errors can quickly lead to failed execution, repeated attempts or unnecessary force applied to sensitive assets. Localised micro-cameras or borescopes are restricted by shadows, debris and optical focal limits during the final millimetres of blind insertion or seal engagement.

This EngD develops an embedded edge-intelligence framework for real-time contact-state estimation, designed to function within hazardous nuclear environments. It is validated on a Lancaster robotic manipulation platform, with TG0's tactile sensing integrated at the end-effector and paired with embedded edge hardware for local tactile inference.

Aims and objectives

1. Develop physically grounded models of tactile signal formation under environmental stressors. Establish how TG0's sensing responds to deformation, contact distribution, temperature variation and potential radiation-induced material drift.
2. Design lightweight, fault-tolerant embedded inference engines. Investigate compact neural models and signal-processing pipelines with fault-detection logic to mitigate transient hardware faults including single-event upsets typical of radiation-exposed COTS embedded systems.
3. Validate embedded contact-state estimation on representative constrained-space tasks. Evaluate estimation accuracy, inference latency, robustness and improvements to right-first-time execution on blind insertion, seal engagement and contact-guided manipulation.

Alignment to STAND-UP impact targets