Article reader Listen + reading controls
Article reader
Preparing the reader…
Reading settings
Disconnected autonomy requires better mission boundaries¶
The Defense Advanced Research Projects Agency (DARPA) is seeking technology for its Rapid Experimental Missionized Autonomy (REMA) program. The objective is to add adaptable autonomy to commercial drones so they can continue a predefined mission when communication with the operator is lost.
Loss of connection is often described as a communications problem. For an autonomous system, it is also an authority problem: what may the machine continue to do when the person can no longer supervise it?
“Continue the mission” is not a complete requirement¶
A human operator interprets intent in context. The operator knows which objective is primary, which constraints are absolute, and which changing conditions justify stopping. A machine needs those boundaries represented explicitly enough to act without inventing authority.
Artificial intelligence (AI)-enabled autonomy should distinguish among:
- the desired outcome;
- permitted actions and geographic limits;
- conditions that require abort, return, hold, or safe landing;
- time and resource limits;
- protected objects and prohibited behavior; and
- what the system should do when its own state estimate is uncertain.
The richer the mission, the more dangerous a vague objective becomes. Optimizing “complete the route” is different from understanding the commander's intent.
Connectivity loss changes the human–machine team¶
When communications are available, the operator can observe, redirect, and resolve ambiguity. Once the link is lost, the level of automation changes suddenly. The system is no longer merely assisting; it is exercising delegated control within the boundaries established beforehand.
Parasuraman, Sheridan, and Wickens's framework for levels of automation helps make this shift visible. Different stages—information acquisition, analysis, decision selection, and action—can be automated to different degrees. A disconnected system may need more authority at several stages, which should trigger different testing and training.
Operators need to know what behavior to expect before the connection fails. They also need a clear handback when communications return. A confused transition of control can be more dangerous than either fully connected or fully disconnected operation.
Graceful degradation should be designed, not hoped for¶
Autonomy will face degraded sensors, uncertain positioning, unfamiliar obstacles, and limited computation in addition to lost communications. The system should reduce ambition as confidence declines.
That may mean selecting a safer route, narrowing the mission, moving to a holding area, or terminating activity. The correct behavior depends on mission and consequence, but it should be specified and tested as part of the design.
Leveson's systems-safety approach focuses attention on control constraints and feedback. For REMA-like systems, designers should identify which feedback channels disappear during disconnection and which local signals can still enforce safe constraints.
Test intent under surprise¶
Evaluation should not only measure whether the vehicle reaches a waypoint. It should introduce situations that force tradeoffs:
- the planned route becomes unsafe;
- location confidence drops;
- a new object enters a protected area;
- remaining energy is insufficient for the original plan;
- sensors disagree; or
- communications return with instructions that conflict with the autonomous state.
The test is whether the system behaves within delegated intent, communicates its uncertainty when possible, and leaves a usable record for review.
The post-mission record closes the loop¶
Disconnected operation should produce a detailed but interpretable trace: what the system perceived, which constraint or objective shaped each important decision, when confidence changed, and why it stopped or continued. Operators and engineers need that trace to distinguish a sound autonomous choice from a fortunate outcome.
The review should update mission templates, training scenarios, and system requirements. Otherwise every unexpected case remains local to the flight.
REMA is pursuing a practical capability for contested and degraded environments. Its success will depend less on giving a drone unlimited independence than on making delegated authority precise. Good autonomy does not act without boundaries when the link disappears. It carries the operator's intent, constraints, and recovery logic into the period when the operator cannot help.
Sources and research trail¶
- Defense Advanced Research Projects Agency, “DARPA Seeks Tech Solutions to Create Autonomous Capabilities for Commercial Drones” (September 12, 2023).
- Department of Defense, Directive 3000.09, Autonomy in Weapon Systems (updated January 25, 2023).
- Parasuraman, Sheridan, and Wickens, “A Model for Types and Levels of Human Interaction With Automation” (2000).
- Leveson, Engineering a Safer World (2011).
- Endsley, “Toward a Theory of Situation Awareness in Dynamic Systems” (1995).