Скачать или смотреть How to make robot understand our language? (feat. VLA)

This clip introduces a systematic study on Vision-Language-Action (VLA) models for generalist robot policies, focusing on optimizing design choices. The authors investigate several critical questions, including which pre-trained Vision-Language Model (VLM) backbone is most suitable, how the VLA architecture should be formulated (favoring continuous action spaces and policy-head history fusion), and the most effective time to leverage cross-embodiment datasets for training. Their extensive experiments lead to the development of a new, highly effective, and open-source VLA framework called RoboVLMs, which achieves state-of-the-art results across simulation and real-world manipulation tasks. The study provides a comprehensive guide for researchers on building robust and generalizable VLA models.