TL;DR TwinVLA(arXiv:2511.05275)는 두 개의 사전 훈련된 단일 팔 VLA를 조합해 양팔 조작(Bimanual Manipulation)을 구현하는 프레임워크다. 양팔 데이터로 처음부터 대규모 사전 훈련 없이, 단일 팔 데이터만으로 사전 훈련된 SingleVLA(0.8B)를 두 개 인스턴스로 구성하고 Joint Attention + Causal Mask로 양팔을 협조시킨다. 결과: RDT-1B(학습 데이터 2,400시간)을 ...
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.
Markdown supported · ⌘↵ to post
No responses yet.