Enabling Physical AI in Trailer Logistics: A Digital Twin Architecture with Continuous Sim-to-Real Feedback

Abstract

Loose-loaded parcels inside 53-foot trailers undergo significant structural rearrangement during highway transit. For autonomous telescopic unloaders, which extract cargo by sweeping planar walls, this transit-induced degradation is catastrophic: a fractured wall cannot be swept efficiently. Solving this requires Physical AI—systems capable of autonomous physical action grounded in real-time sensor data and physics-based reasoning. However, deploying Physical AI in logistics is exceptionally challenging because it requires real-time synthesis of complex states (robotic kinematics, environmental perception) and a decision engine powered by historical knowledge. While robotics systems could attempt to build these representations independently, it is highly impractical. Instead, we propose leveraging the Digital Twin architecture that industries are already rapidly adopting, particularly on the Operational Technology (OT) side. This extended abstract proposes a placement-coordinate formulation hosted within a digital twin. The twin ingests live data from the loading environment (robot joint states, SAM-based RGB-D parcel segmentation, and ML-categorized item types) to make placement decisions. Crucially, we introduce a digital shadow feedback loop: transit IMU data and post-transit unloading assessments are fed back to an offline model to continuously train the policy. By minimizing longitudinal slip, lateral tipping, and sweepability variance, this physics-informed intelligence module enables a digital twin to make loading robots actively aware of downstream unloading constraints.