GCRI · Artificial Intelligence and Autonomous Systems
Autonomous physical systems
Autonomous vehicles, drones, industrial and service robots sense their surroundings and act in the physical world, where errors can injure people and damage property. Their safety depends on perception, decision-making, fallback behaviour and the environments they are allowed to operate in. This page sets out how autonomous physical systems fail, how their safety is assessed, and what evidence serves the regulators, operators and communities involved.
Autonomous physical systems on the record.
Autonomous physical systems: how the risk develops.
Robots, sensors and controls. The dossier below sets out how the risk develops, what is measured, where the data come from, how it is modelled and who decides.
Perception meets the unexpected
Sensors and models misread rare situations: unusual objects, poor weather, glare or confusing road layouts. Rare cases dominate risk because they are hard to test.
Operating domains set limits
Systems are designed for an operational design domain of roads, weather, speeds or spaces. Failures increase when systems are used outside it.
Handover to humans is hard
Supervisors lose attention when automation works most of the time, and cannot take over instantly when it fails.
Safety cases must be argued
Showing that a system is acceptably safe needs a structured argument combining design, testing, simulation and field data, because driving or operating enough to see rare failures is impractical.
Agent oversight console.
The screen a team would use to follow this subject: its measures, states and decisions. Names are examples and values are illustrative.
Agent oversight console
Autonomous physical systems · Agent sandbox · Session AG-7276
Tool calls
- 10:27Infosearch(documents) · read-only
- 10:20Inforun_model(flood_nowcast) · sandbox
- 10:39Gatesend_message(external) · needs approval
- 10:45Errorwrite_file(outside workspace) · denied
Policy
- Scope
- workspace only
- External actions
- human approval
- Spend limit
- 11 units
- Kill switch
- Armed
Actions by approval class
Autonomous machines operate in physical domains with physical risks.
Findings are designed for the Labs Evidence series. SAE J3016 defines levels of driving automation, and the EU Machinery Regulation (2023) addresses machines with self-evolving behaviour.
- Field safety incident data.
- Test protocols for autonomy in unstructured settings.
How autonomous physical systems is modelled, measured and tested.
The model, its assumptions, its data and its uncertainty, in the form a scientist or engineer would check. Structures follow published methods. The values drawn are illustrative, not findings.
System model
The couplings between the systems in autonomous physical systems: signs of influence, a reinforcing and a balancing loop, and cross-links where stress jumps.
What is measured, and how well.
Each indicator carries its unit, its cadence, the kind of source behind it and the evidence label a value would carry. No values are shown: a programme supplies them under its own authority.
| Indicator | Unit | Cadence | Source type | Label |
|---|---|---|---|---|
| Crashes involving automated driving systems | count by severity | monthly | Mandatory crash reporting | Published statistics |
| Disengagements or interventions | count per distance or hour | annual | Operator reports to regulators | REPORTED |
| Distance driven autonomously | kilometres | annual | Operator reports | DECLARED |
| Industrial robot injuries | count | annual | Occupational injury statistics | Published statistics |
| Operating domain exceedances | count of events outside design domain | per deployment | Operator logs | OBSERVED |
| Drone incidents near aircraft | count | annual | Aviation safety reporting | REPORTED |
Where the evidence comes from.
Named public datasets and institutional records, each with its licence class. Sources by agreement are used only under their terms and never republished.
| Source | What it gives | Licence class |
|---|---|---|
| NHTSA Standing General Order crash reports | Crashes involving automated driving and driver assistance systems in the US | Public domain, US government |
| California DMV autonomous vehicle reports | Collision and disengagement reports from permit holders | Public records |
| UNECE regulations on automated driving, including UN Regulation No. 157 | Type approval requirements for automated lane keeping and related systems | Public, UNECE terms |
| ISO 10218 and ISO/TS 15066 | Safety requirements for industrial and collaborative robots | Licensed standards, purchase required |
| National occupational injury statistics | Injuries involving machinery and robots | Public records, licence per government |
Who decides, and what the evidence serves.
Artificial Intelligence and Autonomous Systems prepares evidence. The decisions below are taken by the authorities named, under their own rules, and never by the platform.
| Who decides | The question | Evidence this page serves |
|---|---|---|
| Transport and aviation regulators | Whether and where to permit autonomous operation | Crash data with exposure; safety case reviews |
| Workplace safety regulators | What safety requirements to set for robots | Injury statistics; standards |
| Operators and manufacturers | How to define and enforce operating domains | Domain documentation; exceedance logs |
Outputs this topic produces
- Safety case review frameworks
- Analyses of crash and intervention reports with exposure stated
- Operational design domain documentation templates
- Evidence notes for regulators
What safety engineers ask.
- Does the system perceive its environment correctly across weather, lighting and unusual objects?
- How does it behave at the edges of its operating conditions, and does it know when it is outside them?
- Can it fail safely, handing over to a person or stopping without creating new hazards?
- How is safety demonstrated when learned components cannot be fully specified?
Where physical autonomy fails.
Perception is the weak link: cameras, lidar and radar each fail differently in rain, glare, dust or darkness, and learned perception models can misclassify rare objects. Planning must handle uncertain predictions of other agents’ behaviour. Control must remain stable as conditions change.
Handover to humans is difficult because people monitoring automation lose situational awareness and may not respond in time. Classical safety engineering assumes components whose behaviour can be fully specified, which learned components are not.
Assurance for systems that move.
- 01
Operational design domain
Defining the conditions a system is designed for and detecting when it leaves them.
- 02
Simulation and scenario testing
Testing across large numbers of simulated scenarios, using robotics and automation and high-performance, cloud and edge computing. Simulation fidelity limits what it can reveal.
- 03
Safety cases
Structured arguments, supported by evidence, that a system is acceptably safe for a use, drawing on AI evaluation and assurance engineering. A safety case is only as strong as its weakest evidence.
- 04
Field monitoring
Collecting incidents and near misses after deployment. Under-reporting is common.
How it is studied, and what the methods cannot say.
- 01Robotics and automationScenario-based testing · Collaborative robot safety · Operating domain monitoringApplied here
- 02AI evaluation and assurance engineeringAssurance cases for AI systems · Structured red teaming · Post-deployment monitoringApplied here
- 03Probability and statisticsUncertainty on benchmark scores · Calibration assessment · Audit samplingApplied here
- 04Artificial intelligence and machine learningDistribution shift analysis · Fine-tuning for a specific use · Interpretability methodsApplied here
Where this page’s references lead.
The bodies and documents this page is read against. Status and editions change; each link goes to the issuing body.
Questions and answers.
What is an operational design domain?
Are self-driving cars safer than human drivers?
What is a safety case?
Are autonomous vehicles safer than human drivers?
Can learned components be safety-assured?
Does the platform test products?
Contribute evidence from this page.
Reports is the Consortium’s quarterly publication. What you know about this subject can become evidence, a case or a correction in the next edition, credited to you and published under CC BY 4.0 where it is your own work.
- EmailGateway GW-GCRISend it from this pageEvidence, a correction, a case, a method, a review, a translation or profile input. It arrives tagged with this page, so the editor who owns it receives it.
- ZenodoOpen depositDeposit it and submit to the communityPublish your own record on Zenodo with your ORCID iD, then submit it to the Reports community there. Curators review it; the DOI stays yours.
- ReportsGCRISee every way to contributeThe contribution types, how each is reviewed, what credit you receive, the licence and what is never accepted.
Never send personal data, sensitive infrastructure detail or anything you have no right to share. A contribution is reviewed before use; sending it confers no authorship or editorial rights, and sponsorship never buys a finding.
Bring a method, a dataset or a question.
Information and enquiries are open today. An enquiry receives a reference and names the next step; applications, payments, compute allocations and registrations open only when their team, terms and review capacity exist.
Legal documents.
Maps, data and figures on this site are shown for context and general information. They are not advice, forecasts or warnings; the limits of the information are set out in the Disclaimers and Data Notices.