Building an End-to-End Physical AI Data Pipeline for Autonomous Driving 3.0 on AWS with NVIDIA | Amazon Web Services

AV 3.0 is an end-to-end Reasoning VLA system that performs perception, reasoning, and action under a single unified policy.

Development is organized into 4 phases and 8 stages.

Vehicles continuously record camera, LiDAR, radar, IMU, and GNSS data; a common setup (e.g., 8 cameras, 2 LiDAR) produces several to tens of TB per vehicle per day.

Recordings are packaged in industry-standard container formats (ROS bags .bag, MCAP .mcap, ASAM MDF4 .mf4); ROS bags are the de facto robotics standard and MCAP is the next‑generation format; ASAM MDF4 targets European OEM instrumentation.

Because 5G uplink (ITU IMT‑2020) is limited to 50 Mbps (~540 GB/day), physical media shipment is practical at fleet scale and AWS Data Transfer Terminal handles delivery; raw recordings are stored in Amazon S3 with PB‑scale capacity and S3 Intelligent‑Tiering can move older recordings to lower‑cost storage; the pipeline covers raw collection → AI video curation → neural 3D scene reconstruction → Reasoning VLA training → closed‑loop simulation validation.