Autonomous systems
Aplark Autonomy
From assisted to supervised autonomy.
Aplark Autonomy names where the platform is going: the assurance of systems that decide, not only systems that perceive. It is a research direction and is described here as one.
The problem
Assuring a decision is harder than assuring a perception
The techniques that make a perception component assurable — bounded domain, enumerable cells, measurable per-cell performance — degrade badly when the component under assessment is choosing a course of action over a time horizon. The state space is not enumerable and the failure modes are compositional.
We are not going to claim this is solved. It is the reason the platform exists and it is the last thing on the roadmap, in that order.
How it works
Where the thinking currently sits
The working position is that supervised autonomy — a system that acts within an envelope with a human accountable for the envelope rather than the action — is assurable with an extension of the current evidence model, and that fully unsupervised autonomy in safety-critical roles is not, yet, by anyone.
Aplark Assure is deliberately built so that it never makes a deployment decision autonomously. That boundary is permanent and is not a limitation we intend to remove.
Capabilities
What it does
Envelope-based assurance
The direction under investigation: assuring the bounds within which an autonomous system may act, rather than each action it might take.
Compositional failure analysis
An open research problem being tracked, not a capability under construction.
Who it is for
- Programmes planning for supervised autonomy in mission-critical roles