Sliding Autonomy

Unlike binary handovers, sliding autonomy provides a gradient of control rather than discrete modes. The system can operate at any point between full manual teleoperation and complete independence. This spectrum allows for proportional delegation, where an agent might handle trajectory execution while a human retains target selection authority. The level of autonomy is a continuous variable, often represented as a percentage or a scalar value, enabling smooth transitions without abrupt context switches.

Autonomy levels shift dynamically based on real-time environmental and task assessments. Key triggers include:

• Uncertainty threshold breach: When perception confidence drops below a calibrated limit, control slides toward the human.
• Operational Design Domain exit: Leaving a mapped or validated area automatically increases human oversight.
• Task complexity spikes: Encountering an unmodeled obstacle or a novel manipulation task prompts a takeover request.
• Operator cognitive load: The system can proactively assume more control if it detects the human supervisor is saturated with other fleet demands.

At intermediate points on the spectrum, human and agent contributions are fused mathematically rather than toggled. Common blending architectures include:

• Input mixing: The final actuator command is a weighted sum of the human's joystick vector and the autonomy's planned vector.
• Null-space arbitration: The autonomy controls non-critical degrees of freedom while the human manages the primary task axis.
• Virtual fixtures: The system generates attractive or repulsive force fields that guide the operator's input without overriding it, combining human intent with machine precision for tasks like surgical robotics or precision docking.

Sliding autonomy serves as a dynamic trust interface between operator and machine. By allowing the human to continuously adjust the autonomy level, the system prevents both overtrust (where the operator fails to monitor a highly autonomous agent) and undertrust (where the operator disables useful automation due to lack of confidence). The interface typically displays the current autonomy level and the agent's confidence score for its active task, enabling the operator to make informed adjustments. Over time, as the system demonstrates reliability, the operator naturally slides the autonomy higher.

When a failure or edge case occurs, sliding autonomy provides a structured de-escalation sequence rather than a sudden disengagement. The typical degradation path is:

1. Confidence decay: The system reduces its autonomy level and signals uncertainty.
2. Recommendation mode: The agent proposes actions but awaits human approval.
3. Consent gateway: High-risk actions are queued pending explicit operator authorization.
4. Full manual takeover: Control transfers completely to the human, with the agent maintaining a minimal risk condition as a fallback. This graduated approach prevents panic-driven errors during emergencies.

Every shift toward human control generates a labeled training example for the autonomy stack. The system logs:

• The sensor context that preceded the autonomy reduction.
• The human's corrective action taken during manual control.
• The outcome after the intervention. This intervention logging pipeline feeds into continuous model improvement, allowing the system to gradually expand its Operational Design Domain and reduce the frequency of takeover requests. Over successive deployments, the sliding autonomy midpoint shifts rightward as the agent learns from human demonstrations.