Shared Autonomy

The fundamental mechanism of Shared Autonomy is the real-time arbitration of control authority between the human operator and the autonomous agent. This is not a simple on/off switch but a continuous or discrete blending of inputs.

• Blending: The system combines human commands (e.g., joystick input) with autonomous suggestions (e.g., a collision-free path) into a single, safe control signal.
• Switching: Authority is discretely transferred based on triggers (e.g., human inactivity, detected error, or explicit user request).
• The allocation policy is governed by a meta-controller that evaluates factors like user competence, task difficulty, and environmental uncertainty.

For the autonomous agent to provide useful assistance, it must infer the human's high-level goal. This moves assistance beyond simply filtering low-level commands.

• Methods: Intent can be inferred from direct signals (e.g., gaze, pointing, sparse waypoints) or learned from demonstrated patterns.
• Example: In assistive feeding, the robot uses computer vision to track the user's eye gaze towards a specific food item on a plate, inferring the desired target for the utensil.
• The system maintains a probability distribution over possible goals, updating it as more evidence is observed, and aligns its assistance accordingly.

Both the human and the robot can initiate actions, make suggestions, or correct the other, creating a fluid dialogue. This is key to effective collaboration.

• Robot Initiative: The autonomous system can propose alternative actions ("I suggest a smoother path here"), take over to prevent an error, or ask clarifying questions ("Are you aiming for the blue block?").
• Human Initiative: The human can override, adjust, or ignore the robot's suggestions, maintaining ultimate authority.
• This requires bidirectional communication channels, often using haptic feedback, visual overlays, or auditory cues to signal the robot's intent and state.

The level and type of autonomy provided are not static; they adapt based on the operational context. The system modulates its involvement by assessing:

• Environmental Complexity: Increases assistance in cluttered, dynamic, or hazardous spaces.
• User State: Detects user fatigue, stress, or inexperience and provides more support.
• Task Criticality: Takes a more conservative, safety-enforcing role during high-risk phases of an operation.
• Performance Metrics: Monitors task progress and error rates to adjust its assistance level, aiming to keep the human in the loop (monitoring), on the loop (supervising), or in command as appropriate.

The autonomous system uses internal models (of the world, the task, and the user) to anticipate future states and provide proactive help, reducing the human's cognitive and physical workload.

• Trajectory Prediction: In teleoperation, the robot predicts the user's intended endpoint from initial motion and assists in reaching it precisely.
• Obstacle Avoidance: Continuously calculates potential collisions and subtly modifies the user's command stream to steer around them, often implemented as Virtual Fixtures.
• Next-Tool Selection: In a surgical or assembly task, the robot prepares and positions the next likely required tool based on the procedural stage.

A well-designed shared autonomy system enables seamless and intuitive transitions between levels of control, preventing mode confusion and maintaining user situational awareness.

• Negotiation Protocols: Transitions can be triggered by the human (via a pedal, voice command, or forceful contact), the robot (due to a fault or performance boundary), or jointly agreed upon.
• Graceful Degradation: If autonomy fails, control reverts to the human in a predictable, non-startling manner.
• Example - Driver Assistance: Adaptive Cruise Control (ACC) manages speed, but the driver can instantly override by pressing the accelerator or brake. The system clearly indicates when it is actively controlling (e.g., a green icon) versus when it is only monitoring (a grey icon).

In robotic-assisted surgery, the surgeon operates master controllers while the robotic slave arms execute movements with tremor filtration and motion scaling. The system provides force feedback (haptics) and enforces virtual fixtures—software-defined boundaries that prevent the tools from moving into critical anatomical structures. This blends the surgeon's expertise and decision-making with the robot's precision and stability, reducing tissue trauma and improving patient outcomes.

Modern vehicles implement shared autonomy through systems like lane-keeping assist and adaptive cruise control. The human driver maintains supervisory control, while the system:

• Continuously shares steering authority to keep the vehicle centered.
• Dynamically adjusts braking and acceleration to maintain a safe following distance.
• Issues escalating alerts if driver attention is not detected during system limits. This creates a continuous negotiation of control, enhancing safety and reducing driver fatigue on long journeys.

On factory floors, collaborative robots (cobots) work side-by-side with humans. In a shared autonomy assembly task:

• The human performs high-dexterity steps (inserting a flexible wire).
• The cobot, using force control, holds a heavy component in precise alignment.
• If the human applies intentional force to guide the cobot (hand-guiding mode), the robot complies, learning the trajectory. The system uses power and force limiting (PFL) to ensure safety during physical collaboration, dynamically adjusting its autonomy based on proximity and task phase.

In hazardous environments (nuclear decommissioning, bomb disposal), operators control robots from a safe distance. Shared autonomy mitigates communication latency and complexity:

• The operator specifies a high-level goal ("grasp the valve").
• The robot's onboard autonomy handles low-level stabilization, compliant grasping, and obstacle avoidance.
• Virtual fixtures prevent the arm from colliding with known structures. This blend allows effective operation despite poor video feeds and delayed commands, keeping the human in the strategic loop while the robot manages precise execution.

Next-generation robotic limbs and exoskeletons use shared autonomy to restore natural movement. Intent recognition from residual muscle signals (EMG) or neural interfaces is paired with robotic assistance:

• The user initiates a movement (e.g., reaching for a cup).
• The device's autonomy completes the movement with smooth, stable trajectory control, compensating for user fatigue or tremor.
• The system adapts its assistance level over time based on user performance, promoting neuroplasticity and recovery in therapeutic settings. Control is continuously shared between user intent and robotic support.

For complex inspection tasks (power lines, cell towers), drone pilots use shared autonomy modes:

• The pilot flies the drone generally toward a structure.
• Activating a tracking mode, the drone's vision system autonomously locks onto and follows a cable or beam, freeing the pilot to focus on camera gimbal control.
• Obstacle avoidance runs continuously in the background, creating a safe envelope. The pilot can instantly override with full manual control. This splits the cognitive load: the human handles mission-level decisions, while the robot manages precise, tedious tracking and collision prevention.

A system design principle enabling the dynamic modification of a robot's level of self-governance. Unlike static modes, it allows for smooth, context-driven transitions between fully autonomous, semi-autonomous, and fully manual control.

• Key Mechanism: Provides a slider or selector for the human operator to explicitly set the autonomy level.
• Contrast with Shared Autonomy: Adjustable Autonomy changes the mode of control, while Shared Autonomy dynamically blends control within a mode.
• Use Case: A drone operator switching from autonomous waypoint navigation to manual joystick control during an emergency.

Software-generated guidance or constraint geometries overlaid on a real workspace in a teleoperation or shared control system. They act as haptic or visual guides to channel user inputs, prevent collisions, or improve precision.

• Types: Guidance fixtures (e.g., a tunnel for a surgical tool path) and Forbidden-region fixtures (e.g., a virtual wall protecting sensitive anatomy).
• Implementation: Often uses force feedback on a haptic master device to push the operator away from forbidden regions or pull toward desired paths.
• Role in Shared Autonomy: A primary method for the autonomous system to exert its influence, subtly correcting or augmenting the human's commanded actions.

The process by which a robotic system infers a human's goals or planned actions from observed signals. This inference is critical for the autonomous partner in a shared autonomy framework to provide appropriate assistance.

• Input Signals: Gaze tracking, gesture recognition, motion prediction, physiological data (e.g., muscle activity via EMG), or task context.
• Algorithmic Approaches: Uses probabilistic models (e.g., Bayesian inference), deep learning, or inverse optimal control to map observations to likely intents.
• Example: A robotic exoskeleton detecting a user's intent to stand up from seated posture by analyzing body sway and muscle pre-activation, then providing proportional assistive torque.

An advanced remote manipulation control scheme where a master device operated by a human sends position/velocity commands to a slave robot and simultaneously receives force feedback from the slave's environment. This creates a sense of telepresence.

• Key Feature: Kinesthetic coupling – the human feels the forces and textures encountered by the remote robot.
• Architecture: Requires high-fidelity, low-latency force-reflection control loops to maintain stability and transparency.
• Relation to Shared Autonomy: Serves as the foundational manual control layer. Shared autonomy algorithms can then superimpose autonomous assistance (like virtual fixtures) on top of this bilateral force channel.

A collaborative robot safety mode defined in the ISO/TS 15066 standard. In PFL, the robot's inherent design and control limit its power and force to biomechanically safe thresholds, preventing injury during unexpected contact with a human.

• Safety Basis: Limits are based on pain and injury thresholds for different body regions (e.g., < 150 N for quasi-static contact on the hand).
• Implementation: Uses joint torque sensors and lightweight, back-drivable mechanics to monitor and cap applied forces.
• Enabler for Shared Autonomy: PFL provides the physical safety foundation that allows humans to work in close proximity to a robot, making dynamic shared control feasible without external cages.

A technique where a robot learns a task policy by observing and mimicking one or more demonstrations from a human teacher. Also known as Programming by Demonstration or Imitation Learning.

• Methods: Includes Behavioral Cloning (directly mapping states to actions) and Inverse Reinforcement Learning (inferring the reward function the human is optimizing).
• Sub-method – Kinesthetic Teaching: The human physically guides the robot's end-effector through the task, recording the trajectory as the demonstration.
• Connection to Shared Autonomy: LfD is a primary method for bootstrapping the autonomous policy used in shared control. The robot learns the task from human demonstrations, then uses that learned model to assist during live execution.