SAATT Nav: a Socially Aware Autonomous Transparent Transportation Navigation Framework for Wheelchairs

SAATT Nav: a Socially Aware Autonomous Transparent Transportation Navigation Framework for Wheelchairs

Overview

This work presents the SAATT Nav (Socially Aware Autonomous Transparent Transportation Navigation) framework, a system for autonomous wheelchair navigation that incorporates social awareness and transparent decision-making. The key components are:

• LLM-Based Hypothesis Generation: An LLM is used to generate hypotheses about pedestrian intents (yield, rush, constant velocity), along with natural language rationales.
• Rollout-Based Scoring: The hypotheses are evaluated through forward simulation to assess safety and progress towards the goal.
• Rule-Based Motion Planner: The highest-scoring hypothesis is used to guide the wheelchair's navigation, adapting speed and trajectory based on predicted pedestrian behavior.

The proposed framework is evaluated against baseline methods (A\*, Social Force Model, and an Ablation without the LLM) across three simulated scenarios of increasing social complexity. The results demonstrate that SAATT Nav outperforms the baselines in safety and social compliance metrics while maintaining comparable efficiency and comfort.

Problem & Context

• Powered wheelchairs reduce physical strain but still impose high cognitive load on users due to environmental perception, navigation, and control.
• Autonomous navigation can help, but navigation in crowded settings introduces distinct difficulties around anticipating pedestrian trajectories, respecting social spaces, and interpreting social context.
• Existing autonomous systems lack social awareness and transparent decision-making, leading to decreased perceived safety and trust.

Methodology

System Overview

1. LLM-Based Hypothesis Generation
   • LLM produces a set of hypotheses assigning pedestrian intents (yield, rush, constant velocity) with natural language rationales.
   • Triggered periodically or when pedestrians get close to avoid reaction latency.
2. Rollout-Based Scoring
   • Each hypothesis is simulated forward to evaluate safety (collision count, min. pedestrian distance) and progress towards goal.
   • Scores combined using weights that prioritize safety over progress.
3. Rule-Based Motion Planner
   • Translates the highest-scoring hypothesis into wheelchair control commands.
   • Modulates speed and steering based on predicted pedestrian intent.
   • Detours around detected social bubbles (conversational groups).

Baseline Methods

• A\*: Global path planner treating pedestrians as static obstacles.
• Social Force Model (SFM): Assigns repulsive/attractive forces to pedestrians.
• Ablation: Same planner as SAATT but without the LLM-based reasoning.

Experimental Setup

• 2D simulation with differential-drive wheelchair, circular pedestrians.
• 3 scenarios: open corridor, conversing pair, and crossing pedestrian.
• 30 randomized layouts per scenario, all methods evaluated on the same layouts.
• Metrics across Safety, Social Compliance, Efficiency, and Comfort.

Results & Interpretation

Safety

• SAATT had the fewest collisions and largest minimum pedestrian distance in the social scenarios, indicating better avoidance of collisions and maintaining personal space.

Social Compliance

• In the conversing pair scenario, SAATT spent the least time intruding into the social bubble compared to the baselines.

Efficiency

• A\* was fastest in travel time, likely by ignoring social considerations.
• SAATT performed similarly to the Ablation and SFM in most efficiency metrics.
• No significant differences in success rate (reaching the goal).

Comfort

• SAATT achieved the lowest mean angular velocity in the social scenarios, indicating smoother heading changes.
• A\* had the lowest mean jerk (smoothest acceleration) across all scenarios.

LLM vs. Ablation

• SAATT outperformed the Ablation in safety and social compliance metrics, demonstrating the benefits of intent-based reasoning.
• Similarities in efficiency and comfort metrics suggest the rule-based planner provides a strong base performance.

Transparency

• SAATT generates natural language explanations of its decisions at each waypoint, potentially improving user trust and understanding.

Limitations & Uncertainties

• Currently treats humans as repellent forces, may want to model attraction for social interaction.
• Could incorporate additional social cues like body language, gaze, and vocal cues.
• User preference feedback could personalize the system and avoid over-reliance.
• Inference optimization needed for real-world edge deployment.
• Broader social dynamics should be evaluated in simulation.

What Comes Next

Despite limitations, this work demonstrates the potential of SAATT's approach to autonomous wheelchair navigation that accounts for social awareness and transparent decision-making.