Autonomous Systems
An autonomous system is a machine or piece of software that senses its surroundings, decides what to do and acts on that decision without a person directing each step — a driverless car, a warehouse robot, an inspection drone, an AI agent. Autonomy comes in degrees, and the degree largely decides who is responsible when the system gets it wrong.
- Also known as
- Autonomous machines, Robotics and autonomous systems, RAS
- Used at
- road vehicles and robotaxis, drones and aircraft, warehouses and factories, agriculture, mining and defence, and increasingly software agents that act on accounts and systems
- Components
- sensors (cameras, radar, lidar), perception models that interpret them, a planning and control layer, and a fallback for situations outside the system's design limits
- Levels
- SAE J3016, first published in 2014, sets six levels of driving automation from 0 (none) to 5 (full); from level 3 up the feature counts as an automated driving system rather than driver assistance
- Liability
- the UK's Automated Vehicles Act 2024 makes an authorised self-driving entity, typically the manufacturer or developer, answerable for how the vehicle drives in self-driving mode, and relieves the user in charge of liability for the manner of driving
- AI rules
- the EU AI Act defines an AI system as one designed to operate "with varying levels of autonomy", and requires high-risk systems to be built so that people can oversee them effectively
How autonomous systems work in practice
The core loop is sense, interpret, plan and act, repeated many times a second. What makes a system autonomous rather than merely automated is that it copes with situations its designers did not script one by one: a pallet in the wrong place, a cyclist swerving, a flooded corner of a field. It does so within an operational design domain — the roads, weather, speeds or sites for which it has been validated — and it must recognise when it is leaving that domain and hand over or stop safely.
Most commercial deployments keep a person somewhere in the loop. A warehouse fleet runs on its own, with remote operators resolving the cases the robots flag; a robotaxi service has remote assistance; an AI agent acts within limits and escalates the rest. The economics turn on two numbers: how many machines one supervisor can oversee, and how often the system needs help per hour, mile or task.
For an investor or acquirer the questions are practical. What is the intervention rate and how has it moved? Which domain is the system validated for, and what does each extension cost? Where does liability sit — operator, manufacturer or customer — under the product, road-traffic or AI rules that apply? What does the safety case rest on: simulation, real-world miles, incident records? A system proven in one city or one warehouse layout has not yet shown that it scales.
Example
A logistics start-up runs 120 autonomous forklifts across four customer warehouses. Each makes about 90 pallet moves a day. Over 18 months, interventions — a remote operator clearing a blocked aisle or an unrecognised load — fell from one per 40 moves to one per 400, and one operator now oversees 30 machines instead of 6. The company charges €2,400 per forklift a month, about 60% of the cost of a staffed forklift over two shifts. A buyer's diligence finds that 70% of the clean record comes from two warehouses with identical layouts; at the newest site interventions still run at one per 150 moves. The buyer values the business on the newest site's economics, not the average, and ties part of the price to intervention rates at the next five deployments.
Common mistakes
- Calling a system autonomous when it only follows a fixed script inside a fenced area. - Quoting average intervention rates that hide the newest, hardest deployments. - Leaving the operational design domain vague, so no one knows when the system should stop. - Assuming liability stays with the user; recent laws such as the UK's move it towards the company behind the system. - Confusing the term with an "autonomous system" in internet routing, a network with its own routing policy and number, which has nothing to do with machine autonomy. - Treating remote supervisors as a temporary cost when the business model depends on their ratio to machines.