Teleoperation and Shared Autonomy

Pure teleoperation puts all the work on the person, and pure autonomy needs the robot to know the goal. Shared autonomy sits between: the robot watches the person's commands, infers what they are trying to do, and adds assistance toward that goal. Assistive arms on wheelchairs use this, because fine control through a joystick is tiring.

Maximum assistance Noise in the person's commands Simulated person wants

Or hold the pointer on the canvas to steer yourself.

Inferring the goal

The robot keeps a probability for each possible goal $g$. A command $\mathbf{u}$ that points toward $g$ makes that goal more likely, following a simple Bayesian update with a rationality parameter $\beta$:

$$P(g \mid \mathbf{u}_{1:t}) \propto P(g)\,\prod_k \exp\!\big(\beta\,\cos\angle(\mathbf{u}_k,\ g - \mathbf{x}_k)\big)$$

Blending

The robot's assistance $\mathbf{u}_r$ points at the most likely goal, and the final command blends the two, with more help when the robot is confident:

$$\mathbf{u} = (1 - \alpha)\,\mathbf{u}_h + \alpha\,\mathbf{u}_r, \qquad \alpha = \alpha_{\max}\cdot\max_g P(g)$$

What to notice

With noisy commands and no assistance, the path wanders. With assistance, once the robot becomes confident it straightens the path toward the right object. Early on, when two objects lie in similar directions, it is unsure and holds back, which matters, because confident help toward the wrong goal is worse than none. Real systems also let the person override instantly and show what the robot thinks they want.