ChikaKomari/26summe-camp-10-data

StarVLA数据集是一个用于视觉语言动作（VLA）训练的多模态机器人数据集集合，包含多个子数据集：CALVIN（基于Panda机器人的任务数据集，包含17870个episode和1071743帧，用于动作预测）、OXE（包括Bridge/WidowX和Fractal/RT-1数据集，分别涉及widowx和google_robot机器人，用于跨实施例训练）、RoboCasa（基于GR1/Fourier Hands的24个任务数据集，每个任务1000个episode，用于桌面操作任务）和RoboTwin 2.0（包含Clean和Randomized任务，涉及aloha机器人，用于双臂操作任务）。这些数据集覆盖了不同机器人平台、任务类型和动作模态，支持多框架训练（如QwenPI、QwenOFT等），用于训练视觉语言动作模型以实现机器人控制。The StarVLA dataset is a collection of multimodal robotic datasets for Vision-Language-Action (VLA) training, comprising several sub-datasets: CALVIN, a task dataset based on the Panda robot containing 17,870 episodes and 1,071,743 frames for action prediction; OXE, which includes the Bridge/WidowX and Fractal/RT-1 datasets involving the widowx and google_robot robots respectively and used for cross-embodiment training; RoboCasa, a dataset of 24 tasks based on GR1/Fourier Hands with 1,000 episodes per task targeting tabletop manipulation tasks; and RoboTwin 2.0, which includes Clean and Randomized tasks involving the Aloha robot and designed for dual-arm manipulation tasks. These datasets cover diverse robotic platforms, task types and action modalities, support training with multiple frameworks (e.g., QwenPI, QwenOFT, etc.), and are used to train Vision-Language-Action models for robotic control.