Shared Autonomy

A control scheme where the robot infers a human operator's intent and blends its own autonomous action with the person's input.

Shared autonomy sits between pure teleoperation and full autonomy: a person gives rough, noisy commands through a joystick, brain-computer interface, or similar, while the robot infers what the person is trying to achieve and blends that inferred goal with the person's own input — commonly by weighting the mix according to how confident it is about the person's intent, an approach called arbitration or policy blending. It was first used for assistive robotic arms, smart wheelchairs, and teleoperation, aiming to reduce the operator's workload without taking away their control. Siddhartha Javdani and colleagues at CMU modeled it in 2015 as a partially observable Markov decision process (POMDP) with an unknown goal, using inverse optimal control to infer a distribution over goals from the person's history of inputs; users completed tasks faster with less input, though some reported feeling they had lost a sense of control. Reddy, Dragan, and Levine used deep reinforcement learning in 2018 to remove the need for a predefined set of possible goals.

ExampleIn an experiment by Reddy and colleagues, a person plays the Lunar Lander game while an assistive agent first filters out actions whose Q-value falls below a threshold, then picks whichever of the remaining actions is closest to the human's input — helping the person land more steadily without ever knowing the actual target landing site.

Also called

Shared Control

Related

Teleoperation · Human-in-the-Loop · Human-Robot Collaboration · Human-Robot Interaction · Intent Understanding · Partially Observable Markov Decision Process

Sources

Shared Autonomy via Hindsight Optimization (Javdani, Srinivasa, Bagnell, arXiv 1503.07619)
Shared Autonomy via Deep Reinforcement Learning (Reddy, Dragan, Levine, RSS 2018, arXiv 1802.01744)