Transformer-based human-motion forecasting coupled with safe reinforcement learning for telepresence robot...

Area of Science:

• Robotics
• Artificial Intelligence
• Human-Robot Interaction

Background:

• Telepresence robots (TPRs) require safe co-navigation with humans in confined hospital settings.
• Current methods often react to human motion instead of anticipating it, limiting robustness in busy areas.
• Integrating short-horizon human-motion forecasting with safety-constrained control is crucial for improving TPR navigation.

Purpose of the Study:

• To evaluate an anticipatory, safety-aware co-navigation framework for TPRs in hospital environments.
• To address the gap in integrating human-motion forecasting with safety-constrained control for enhanced robot robustness.
• To improve the safety and reliability of TPRs in human-dense clinical layouts.

Main Methods:

• Developed a modular framework combining a transformer-based forecaster for multi-agent trajectory prediction with a safe reinforcement learning (RL) controller.
• Integrated predicted pedestrian states into the RL policy as risk-aware occupancy features.
• Enforced safety using constrained policy optimization and a control barrier function (CBF) shield to filter unsafe actions.
• Benchmarked against social-force/DWA, attention-based crowd-RL, and MPC+CBF across hospital-like corridor and ward benchmarks.

Main Results:

• The proposed method improved task success by 21.6% and reduced collisions by 47.3% compared to the best baseline.
• Median minimum human-robot clearance increased by 0.19 m, and near-miss events decreased by 38.5%.
• Time-to-goal was comparable to MPC+CBF with zero CBF violations, while ablations showed forecasting and CBF shield were critical for performance.

Conclusions:

• Anticipatory perception coupled with Safe-RL significantly enhances telepresence co-navigation safety and reliability in hospitals without compromising efficiency.
• The modular framework allows for flexibility in choosing alternative forecasters and safety shields.
• Future work includes exploring on-device adaptation and incorporating operator intent for live hospital workflow evaluations.