TwinVLA: Bimanual Manipulation from Two Single-Arm VLAs — Outperforming RDT-1B with 50 Episodes

I build robots for a living. Not in simulation. Not in a lab. On the floor, with real hardware, real failure modes, and real deadlines. My work spans the full stack of modern robotics: embedded systems firmware, autonomous navigation, and the end-to-end pipeline for Vision-Language-Action (VLA) model development — dataset collection, training, inference optimization, and sim-to-real transfer. I've worked hands-on with platforms from Unitree, Deep Robotics, Dexmate, and Robotis, among others. Each one teaches you something different about the gap between what models promise and what robots actually do.