Lightning Talk: Full-Stack PyTorch Robotics VLA: From Data To... Samet Akcay & Dmitriy Pastushenkov

This talk covers the transition of Vision-Language-Action models to production in robotics, addressing the challenges of fragmented edge deployment paths. The speakers, Samet Akcay and Dmitriy Pastushenkov from Intel, present a unified, PyTorch-native workflow that encompasses the full robotics lifecycle, including data capture, curation, and optimized edge execution. They introduce a modular Physical AI pipeline to bridge the gap between research scripts and real-time hardware, detailing practical patterns for integrating robotics data capture and policy training in a PyTorch ecosystem. The session also outlines how to export models using ExecuTorch with an OpenVINO backend, tackling issues related to latency, accuracy, and operator coverage. Attendees will gain insights into a reference architecture, monitoring strategies, and best practices for managing dataset drift, making it easier to deploy robotics VLA systems in production environments.