Effects of Robot Competency and Motion Legibility on Human Correction Feedback

TL;DR

This work investigates how robot competency and motion legibility influence how humans supervise and correct robotic behavior in a Learning from Corrections (LfC) setting. Using a between-subject study (N=$60$) with $64$ pick-and-place trials per participant, the authors analyze correction timing, necessity, and the precision-effort trade-off under manipulated competency and legibility. They find that legible and predictable motions intensify sensitivity to high competence, with both under- and over-correction patterns depending on competency, and they empirically validate a general positive relation between correction precision and physical effort, though this relation weakens for incompetent robots with legible motions. The results have practical implications for designing robot interactions and for developing learning algorithms that interpret supervisory feedback in deployment contexts.

Abstract

As robot deployments become more commonplace, people are likely to take on the role of supervising robots (i.e., correcting their mistakes) rather than directly teaching them. Prior works on Learning from Corrections (LfC) have relied on three key assumptions to interpret human feedback: (1) people correct the robot only when there is significant task objective divergence; (2) people can accurately predict if a correction is necessary; and (3) people trade off precision and physical effort when giving corrections. In this work, we study how two key factors (robot competency and motion legibility) affect how people provide correction feedback and their implications on these existing assumptions. We conduct a user study ($N=60$) under an LfC setting where participants supervise and correct a robot performing pick-and-place tasks. We find that people are more sensitive to suboptimal behavior by a highly competent robot compared to an incompetent robot when the motions are legible ($p=0.0015$) and predictable ($p=0.0055$). In addition, people also tend to withhold necessary corrections ($p < 0.0001$) when supervising an incompetent robot and are more prone to offering unnecessary ones ($p = 0.0171$) when supervising a highly competent robot. We also find that physical effort positively correlates with correction precision, providing empirical evidence to support this common assumption. We also find that this correlation is significantly weaker for an incompetent robot with legible motions than an incompetent robot with predictable motions ($p = 0.0075$). Our findings offer insights for accounting for competency and legibility when designing robot interaction behaviors and learning task objectives from corrections.

Figures (11)

• Figure 1: The robot begins moving along the yellow trajectory to place a shape in one of the target holes. The participant intervenes, guiding the robot to a different target via the green trajectory, leaving the dashed portion of the robot's intended trajectory untraveled.
• Figure 2: A comparison between predictable (solid lines) and legible motions (dashed lines) starting from S to two potential goals, A and B.
• Figure 3: This figure depicts the series of pick-and-place tasks performed by the robot. The tasks involved manipulating various shapes in sequence: circle, square, triangle, and rectangle. Each shape had 4 colors: orange, red, green, and blue. The colored lines illustrate the robot's trajectories, categorized by their legibility: cyan represents legible paths, yellow represents predictable paths, and purple represents illegible paths.
• Figure 4: In an intended-success trial, the robot will attempt to place a red square shape into the red target. In an intended-failure trial, the robot will attempt to place the red shape into the wrong-colored target or onto the workbench surface. The yellow-shaded squares represent the distribution of possible locations where the robot may attempt to place the shape.
• Figure 5: Task Objective Divergence: The KLD between the estimated robot's task belief distribution and the actual, correct motion goal.