Adjustable Autonomy

The system's defining feature is the ability to smoothly transfer control authority between human and robot along a spectrum. This is not a simple on/off switch but a continuum, often visualized as levels (e.g., from fully manual teleoperation to supervised autonomy to fully autonomous operation). Transitions can be initiated by the human (e.g., pressing an e-stop or taking the wheel), triggered by the robot (e.g., a 'hand-off' request when it encounters uncertainty), or enforced by the system (e.g., an automatic takeover in a safety-critical situation).

Adjustable autonomy systems use contextual awareness to recommend or enact appropriate autonomy levels. This decision is based on a real-time fusion of factors:

• Environmental Complexity: Cluttered vs. structured space.
• Task Criticality: Consequences of failure (e.g., surgical procedure vs. warehouse sorting).
• Human State: Operator workload, attention, or explicit intent.
• Robot Confidence: The system's own estimated probability of success given sensor uncertainty. For example, an autonomous forklift might switch to human-guided mode when navigating a densely packed, dynamic aisle it has not mapped before.

Adjustment can occur at different scopes of control, which is a key design consideration.

• Full-System Adjustment: The entire robotic platform's autonomy level is changed (e.g., switching a drone from autonomous flight to manual joystick control).
• Subsystem or Task-Level Adjustment: Specific functions are adjusted independently. For instance, a surgical robot's navigation might be autonomous while its delicate cutting action remains under direct surgeon control.
• Mixed-Initiative Interaction: The human and robot control different aspects simultaneously in a shared autonomy paradigm, such as a human specifying a high-level goal while the robot handles low-level trajectory optimization and obstacle avoidance.

Effective adjustable autonomy requires intuitive human-machine interfaces (HMIs) that make the current autonomy mode, available modes, and transition mechanisms clear. This includes:

• Mode Awareness: Visual, auditory, or haptic signals indicating 'who is in charge'.
• Transition Triggers: Physical buttons, voice commands, or gesture controls for the human to request a change.
• Robot-Initiated Requests: The system must communicate its need for help or its intent to take over in a way that doesn't startle or confuse the operator. This is critical for trust calibration.

All autonomy adjustments are bounded by a safety architecture. The system must ensure safe states during and after any transition. This involves:

• Fail-Safe Defaults: Defaulting to a safe stop or low autonomy level on system fault.
• Compliance with Standards: Adherence to frameworks like ISO/TS 15066 for collaborative robots, which defines safety-rated monitored stops and power and force limiting (PFL).
• Predictable Behavior: Transitions must be deterministic and free from dangerous latent states to prevent accidents caused by mode confusion.

For smooth collaboration, both human and robot should operate on a shared situational understanding. The robot maintains a world model it can visualize for the operator, highlighting its plans, perceived obstacles, and uncertainties. Conversely, the system employs intent recognition algorithms to infer human goals from gaze, gesture, or motion, allowing it to act proactively. This bidirectional understanding reduces the cognitive load of monitoring and makes control transfers feel natural rather than disruptive.

In robotic-assisted surgery, adjustable autonomy enables critical transitions between modes. The surgeon may perform a delicate incision under direct teleoperation, then command the system to autonomously hold a steady position or suture along a pre-planned path. This dynamic handoff optimizes precision, reduces surgeon fatigue, and maintains ultimate human oversight for life-critical decisions. Systems like the da Vinci Surgical System exemplify this principle, though full autonomy in surgery remains a carefully guarded frontier.

This domain features clear, multi-level autonomy spectra. A mining truck may operate fully autonomously in a geofenced pit but require remote teleoperation for loading. For consumer vehicles (SAE Levels 0-5), adjustable autonomy manifests as:

• Handover Requests: The system detecting an edge case and requesting driver takeover.
• Driver Monitoring: Systems ensuring the human is ready to resume control.
• Fallback Modes: Graceful degradation to a safer minimal-risk condition if the human does not respond. The transition protocol is a core safety-critical challenge.

Collaborative robots (cobots) are prime examples of adjustable autonomy in structured environments. Modes defined by standards like ISO/TS 15066 include:

• Hand Guiding: The robot is fully compliant, moved by the worker to teach a path.
• Power and Force Limiting (PFL): The robot operates autonomously but with capped speed and force for safe contact.
• Safety-Rated Monitored Stop: Full stop on human approach, automatic restart on exit. This allows a single robot to switch between autonomous assembly, human-assisted kitting, and being manually positioned for new tasks.

In unstructured, hazardous environments, adjustable autonomy is vital for resilience. A ground robot exploring rubble may autonomously navigate but switch to teleoperation for complex obstacle negotiation. Key patterns include:

• Supervised Autonomy: The operator sets high-level waypoints; the robot autonomously plans and executes paths between them.
• Mixed-Initiative Control: The robot suggests actions (e.g., "I can enter that crevice") for the operator to approve.
• Contextual Triggers: Autonomy level automatically reduces when communication latency increases or sensor confidence drops.

Extreme communication latency (minutes to hours) makes direct teleoperation impossible, necessitating high autonomy. Adjustable autonomy here operates on long timescales and abstract commands.

• Ground-in-the-Loop: Scientists send a high-level goal ("sample rock at location Alpha"). The rover autonomously generates and executes a day-long activity plan, reporting back.
• Exception Handling: If an anomaly occurs, the system may enter a safe, stationary mode and await new instructions, or execute a pre-defined contingency plan. This paradigm, used by NASA's Mars rovers, maximizes scientific return despite the delay.

For robots in human homes, autonomy must adapt to user ability and preference. A mobile manipulator assisting with daily living might:

• Offer full autonomy for routine tasks like fetching a known item.
• Shift to shared control for delicate tasks like pouring a drink, where the user guides the arm's gross motion while the robot stabilizes the wrist.
• Accept direct joystick teleoperation for entirely novel situations. The adjustment can be user-initiated ("Robot, let me drive") or system-initiated based on perceived user struggle or task uncertainty, requiring robust intent recognition.

A control paradigm where authority over a task is dynamically blended between a human operator and an autonomous system. Unlike a simple mode switch, it continuously fuses human inputs (e.g., joystick commands) with machine-generated plans to produce a single, collaborative control signal. This is the algorithmic engine that often enables smooth Adjustable Autonomy.

• Key Mechanism: Uses techniques like arbitration filters or potential fields to combine inputs.
• Example: A surgical robot dampens a surgeon's hand tremor while also providing virtual fixtures to prevent instrument collision.

A robot designed with inherent safety features to operate alongside humans in a shared workspace without traditional safety cages. Cobots are the primary hardware platform for implementing Adjustable Autonomy in industrial settings.

• Core Features: Force/torque sensing in joints, rounded edges, low inertia, and compliant actuators.
• Operation Modes: Defined by ISO/TS 15066, including Hand Guiding, Power and Force Limiting (PFL), and Safety-Rated Monitored Stop—all forms of adjustable autonomy.
• Example: A Universal Robots UR10 arm can be manually positioned by an operator (hand guiding) and then execute the recorded trajectory autonomously.

A collaborative operation safety mode defined in ISO/TS 15066 where a robot's design intrinsically limits its power and force to biomechanically safe thresholds. It is a pre-requisite safety layer that makes physical Adjustable Autonomy possible.

• Biomechanical Limits: Defines maximum permissible values for transient and quasi-static contact pressure and force on 29 body regions.
• Implementation: Requires integrated joint torque sensors and control algorithms that enforce strict dynamic limits.
• Purpose: Ensures that unintended contact during collaborative tasks (like hand guiding or co-manipulation) will not cause injury.

The process by which a robotic system infers a human's goals or planned actions from observed signals. It is a critical enabling technology for proactive Adjustable Autonomy, allowing a system to anticipate needed assistance or mode changes.

• Input Modalities: Gaze tracking, gesture recognition, motion trajectory analysis, physiological sensors (EMG, EEG), or natural language.
• Application: A mobile robot observes a human reaching for a heavy box and autonomously moves to provide support, transitioning from a passive to an assistive autonomy level.
• Challenge: Requires robust probabilistic models to avoid misinterpretation and inappropriate autonomous action.

A master-slave control architecture for remote manipulation where the human operator receives kinesthetic force feedback from the robot's environment. It represents one extreme of the Adjustable Autonomy spectrum: direct, continuous human control augmented by machine-mediated transparency.

• Key Feature: The 'bilateral' nature—command signals flow to the slave robot, and force signals flow back to the master device.
• Enables: A sense of telepresence, allowing delicate tasks like underwater manipulation or bomb disposal.
• Autonomy Integration: Often enhanced with Virtual Fixtures (software constraints) that provide guidance or prevent collisions, blending direct control with autonomous safeguarding.

The process of aligning a human user's trust in a robot's capabilities with the robot's actual performance. It is a critical human-factors challenge for effective Adjustable Autonomy, as inappropriate trust (over or under) degrades system performance and safety.

• Over-Trust: Leads to complacency and lack of supervision when the robot is error-prone.
• Under-Trust: Leads to inefficient micromanagement and rejection of useful autonomous functions.
• Calibration Methods: Use of Explainable AI (XAI) interfaces, performance history displays, and predictable robot behavior to build appropriate mental models.