We study transparency as an interface design variable, defined by feedback modality (visual vs. auditory) and information richness (sparse vs. rich). We implement our interfaces on VOSA, a vision-based shared autonomy framework that performs real-time intent inference from egocentric visual observations without task-specific training.

Visual Sparse highlights the single object that the robot currently infers as the user's intended target.

Visual Rich displays bounding boxes, labels, and confidence scores for all candidate objects, with color encoding confidence from red (low) to green (high). 

Auditory Sparse verbalizes only the currently inferred object label when the robot's prediction changes.

Auditory Rich verbalizes label and confidence information for all candidate objects that meet the confidence threshold.

We evaluate these interfaces across two assistive manipulation tasks differing in complexity.

Shelving: placing bottles and condiments onto a two-level shelf. 

Sorting (Recycling): sorting cluttered objects into the correct recycling bins, a more complex scenario due to increased clutter and a larger candidate set. 

We conducted a user study (N = 25) across both tasks, with participants teleoperating a Kinova Gen3 manipulator via an Xbox controller.

Providing feedback significantly improves intent alignment and reduces the need for corrective interventions, indicating that making the inferred goal legible accelerates convergence in shared control. This holds regardless of feedback modality or richness.

Revealing the full belief distribution did not consistently improve alignment or trust. Users did not require full access to the robot's internal belief distribution to realize these coordination benefits; exposing the inferred goal alone was sufficient.

Participants preferred visual feedback over auditory feedback, reporting greater understanding of the robot's plan and higher trust when intent information was presented visually.

Preferences for sparse versus rich information depended on task complexity. In the shelving task, participants preferred Visual Rich (52%). In the more cluttered and ambiguous sorting task, participants preferred Visual Sparse (48%), reflecting a trade-off between completeness and cognitive load.

Auditory feedback was most tolerable in its sparse form. Rich auditory feedback was the least preferred condition overall, with participants describing it as distracting rather than informative.

Based on these results, we outline the following guidelines for designing transparent shared autonomy systems.

Transparency design should emphasize goal legibility over belief disclosure. Exposing the inferred goal is sufficient to support coordination; transparency need not be exhaustive to be effective.

Favor visual feedback for continuous spatial interaction. Visual feedback aligns with the spatial, continuous nature of assistive manipulation tasks and can be referenced without interrupting ongoing control.

Match feedback richness to task complexity and environmental ambiguity. Richer feedback supports understanding in simpler environments but becomes counterproductive in cluttered, ambiguous ones.

Use auditory feedback selectively and sparingly. Auditory transparency is most effective when limited to brief cues indicating salient intent changes rather than sustained verbal explanation.