Do Robots Need Body Language? Comparing Communication Modalities for Legible Motion Intent in Human-Shared Spaces

1. Problem Statement

As legged robots (e.g., quadrupeds) increasingly operate in human-shared spaces (sidewalks, crosswalks), there is a critical need for them to communicate their motion intent clearly. Currently, robots often move in ways that are difficult for humans to interpret, forcing pedestrians to adapt and increasing the risk of startling or collision.

• The Core Challenge: While prior work has optimized trajectories for "legibility" (making goals inferable from path), it remains unclear how implicit cues (expressive body language/motion) compare to explicit channels (lights, text, audio) in conveying specific navigation actions.
• The Gap: There is limited empirical data on how these different modalities perform individually, how they interact when aligned (redundant), and how they affect user prediction accuracy, confidence, and trust when they conflict.

2. Methodology

The authors conducted an online video-based study using the Boston Dynamics Spot quadruped robot.

• Participants: 210 participants recruited via Prolific (demographically diverse; mean age ~42).
• Experimental Design:
  • Scenarios: Four common navigation contexts: Crosswalk, Turning, Passing, and Starting Movement.
  • Signal Modalities: Four channels were tested:
    1. Implicit: Expressive body motion (directional leaning, limb gestures, deferential postures like crouching/bowing).
    2. Explicit 1: Indicator lights (LEDs).
    3. Explicit 2: Text displays.
    4. Explicit 3: Synthesized audio.
  • Conditions: The study utilized a combinatorial design to create 78 unique video clips categorized into:
    • No Signal: Baseline.
    • Single Channel: One modality active.
    • Redundant: Multiple modalities aligned (e.g., Body Motion + Text saying "Turn Left").
    • Conflicting: Modalities expressing opposing intents (e.g., Body leaning Left + Text saying "Turn Right").
• Procedure: Participants viewed 10 randomized clips and answered three questions after each:
  1. Prediction: What will the robot do next? (Accuracy metric).
  2. Confidence: How sure are you? (1–7 Likert scale).
  3. Trust: How much do you trust the robot to move safely? (1–7 Likert scale).

3. Key Contributions

1. Combinatorial Experimental Design: A systematic comparison of implicit (body language) vs. explicit (lights, text, audio) modalities on a single robot platform across identical tasks.
2. Analysis of Signal Alignment: Empirical evidence on how redundancy (aligned signals) and conflict (misaligned signals) impact human perception, moving beyond single-modality studies.
3. Quantitative Metrics: A comprehensive dataset measuring the trade-offs between prediction accuracy, user confidence, and trust across different signaling strategies.

4. Key Results

The study yielded distinct findings regarding the effectiveness of different modalities:

• Prediction Accuracy:
  • Baseline (No Signal): Lowest accuracy (~14%), indicating humans struggle to infer intent from raw trajectory alone.
  • Implicit (Body Motion): Significantly improved accuracy to ~44%. This demonstrates that even non-optimized, inherent body language provides meaningful information.
  • Explicit Channels: Performed best. Text (~88%) and Audio (~82%) were the most accurate. Lights (~58%) were moderate.
  • Redundancy: Adding redundant signals (e.g., Body + Text) did not significantly increase accuracy over the single best explicit channel (Text/Audio).
• Confidence:
  • Explicit channels (Text/Audio) and Redundant conditions generated the highest confidence (Mean > 5.5).
  • Interestingly, Body Motion slightly reduced confidence compared to the no-signal baseline (Mean ~4.2 vs 4.7), suggesting that while informative, body cues were perceived as less reliable than explicit text/audio.
  • Conflicts: Misaligned signals caused a sharp drop in confidence, particularly when implicit motion conflicted with explicit text/audio.
• Trust:
  • Trust ratings followed a similar trend to confidence but showed no statistically significant differences between most conditions.
  • No-signal conditions had the lowest trust (~4.5). Text and Audio achieved the highest trust (>5.0).
  • Conflicts caused a slight decrease in trust, but less than anticipated.

5. Significance and Conclusion

• Validation of Body Language: The study confirms that expressive body motion is a viable, low-cost communication strategy. Even without optimization, it nearly triples prediction accuracy compared to no signaling. This supports the design of robots that use inherent morphology (leaning, gait) to communicate intent without requiring extra hardware.
• The Primacy of Explicit Channels: Text and Audio remain the "gold standard" for clarity and confidence. However, the authors note practical limitations (language barriers, noise, visual clutter) where these may fail.
• Redundancy vs. Cognitive Load: While aligned signals increased confidence slightly, they did not improve accuracy, suggesting a "ceiling effect" where adding more channels may introduce cognitive load without adding informational value.
• Conflict Management: The finding that conflicts degrade confidence but not necessarily trust suggests that users may remain trusting of a robot even when its signals are confusing, provided the robot does not act dangerously.

Final Takeaway: Robots in human-shared spaces should prioritize explicit channels (text/audio) for maximum clarity but must integrate intentional expressive motion as a foundational layer of communication. Body language is essential for scenarios where explicit channels are unavailable or ambiguous, serving as a critical "safety net" for legibility.