Legible Shared Autonomy: Implicit Communication of Robot Belief through Motion

Technical Summary: Legible Shared Autonomy

Problem Statement
Shared autonomy systems assist users with motor impairments by combining user input with autonomous assistance to control robot arms. However, a fundamental "transparency problem" limits their effectiveness: users cannot observe the robot's internal beliefs regarding the user's intended goal. Traditional systems prioritize efficiency, moving along the shortest path to the inferred goal. When multiple objects lie in similar directions (e.g., a tissue and a cup on a cluttered table), this efficient motion is ambiguous. This creates two critical failure modes:

1. Unnecessary Effort: When the robot correctly infers the goal, users continue to provide control input because they cannot perceive that the robot understands their intent, wasting effort that could be delegated to the robot.
2. Delayed Error Correction: When the robot misunderstands the intent, users cannot detect the error until the assistive motion diverges significantly from their desired trajectory, requiring substantial corrective input.

Methodology
The paper proposes a Legible Shared Autonomy framework where the robot communicates its inferred beliefs about user goals through the characteristics of its assistive motion, rather than relying on explicit feedback interfaces.

• Action-Level Legibility Metric: Unlike prior trajectory-level approaches optimized offline, this work introduces an action-level metric computable at each timestep for real-time interaction. The metric quantifies how well a robot action $a_R$ discriminates the inferred goal $\theta^*$ from competing alternatives $\theta$. It is defined as the log-ratio of the likelihood of the action under the inferred goal versus the most confusable alternative, based on a Boltzmann-rational user action model:
  $$L(a_R | s_t, \theta^*) = \log \pi_H(a_R | s_t, \theta^*) - \log \max_{\theta \neq \theta^*} \pi_H(a_R | s_t, \theta)$$
  Positive values indicate the action strongly signals the inferred goal; values near zero indicate ambiguity.

• Dual Optimization Framework: The robot selects actions by balancing task performance (minimizing distance to the goal) and legibility. The objective function is:
  $$a^*_R = \arg \max_{a_R \in A_R} [\lambda L(a_R | s_t, \theta^*) + Q(a_R, s_t, \theta^*)]$$
  where $\lambda$ weights the importance of legibility. When $\lambda=0$, the system reverts to standard efficient shared autonomy.

• Confidence-Aware Adaptive Authority Allocation: To manage the trade-off between revealing beliefs and maintaining control, the system dynamically adjusts the authority weight $\beta_t$ given to the user. This weight is modulated by two factors:

  1. Inference Confidence ($b_t$): The robot grants itself more authority (lower $\beta_t$) only when its belief confidence exceeds a threshold.
  2. Proximity to Goal ($d_t$): The robot increases its authority as it nears the goal, where precise goal discrimination is most critical.
     When the user releases the input device, the robot assumes full control, allowing delegation once the user perceives correct inference.

Key Contributions

1. Action-Level Legibility Metric: A computationally efficient metric enabling real-time legible intent communication, distinct from offline trajectory optimization.
2. Legible Shared Autonomy Framework: The first integration of legibility into shared autonomy to reveal beliefs about user goals (rather than the robot's own goals), coupled with an adaptive authority strategy that signals confidence levels.
3. Empirical Validation: Comprehensive user studies demonstrating that legible motion improves transparency and reduces user effort.

Experimental Results
The approach was validated through two studies: a 2D simulation with 20 participants and a physical experiment with a 6-DOF Dobot CR5 robot arm with 15 participants.

• Transparency and Understanding: In the simulation, legible motion significantly improved users' ability to perceive whether the robot understood their intent (increasing from 38% in standard SA to 96–99% in legible SA) and correctly identify the inferred goal.
• Subjective Experience: Physical experiments showed a 62% increase in perceived intuitiveness and a 55% increase in collaboration quality scores for legible shared autonomy compared to standard shared autonomy. A strong positive correlation ($r=0.926$) was found between perceived understanding and collaboration quality.
• Control Effort: Legible shared autonomy significantly reduced user control effort by 27% in physical experiments (and 47% in simulation), as users could delegate control earlier when they perceived correct inference.
• Task Duration Trade-off: The study noted a 12% increase in task duration in physical experiments. This is attributed to trajectories deviating from the shortest path to establish clear goal identity. The authors argue this trade-off is acceptable for assistive robotics, where reducing physical workload and increasing system comprehensibility are prioritized over raw speed.

Significance and Claims
The paper claims that legible shared autonomy transforms the interaction from one-way inference into transparent bidirectional collaboration. By modulating motion characteristics to reveal internal beliefs, the system allows users to dynamically adjust their control strategies—reducing input when understanding is confirmed or correcting errors immediately when misunderstandings are detected.

The authors conclude that while legibility should be tuned to the task rather than indiscriminately maximized (as overly assertive motion can sometimes reduce perceived collaboration), assistive motion can effectively carry information about the robot's belief state. This bridges the transparency gap, moving shared autonomy from "silent assistance" toward more comprehensible human-robot teaming. The work acknowledges limitations, including the need to evaluate high-confidence incorrect inferences and generalization to cluttered scenes with many candidate goals.