Keh0t0/dreamgen_robocasa_p3_init0_v2_droid

DreamGen Robocasa P3 (droid-IDM) — dreamgen_robocasa_p3_init0_v2_droid是一个合成的LeRobot v3.0数据集，通过Cosmos-Predict2 14B自回归视频生成技术创建，并采用新的droid-IDM动作和状态推断方法。数据集包含1440个片段，覆盖24个任务，每个任务有60个条目。数据模式包括三个摄像头视图（robot0_agentview_left/right_image和robot0_eye_in_hand_image，分辨率为128×128，经过中心裁剪和h264编码），动作数据为12维浮点数组（h50 action_eef空间），状态数据为16维浮点数组（包括末端执行器位置、四元数、夹爪位置和基座位置等）。生成流程涉及从Cosmos AR视频到LeRobot v3.0构建，再到droid IDM处理，以输出动作和状态。数据集设计用于与真实数据集（如Keh0t0/robocasa_mg30_real_v2）联合训练，以提升机器人学习性能。DreamGen Robocasa P3 (droid-IDM) — dreamgen_robocasa_p3_init0_v2_droid is a synthetic LeRobot v3.0 dataset generated via Cosmos-Predict2 14B autoregressive video generation with new droid-IDM action and state inference. The dataset consists of 1440 episodes covering 24 tasks with 60 entries each. The schema includes three camera views (robot0_agentview_left/right_image and robot0_eye_in_hand_image, at 128×128 resolution, center-cropped and h264-encoded), action data as a 12-dimensional float32 array (h50 action_eef space), and observation state as a 16-dimensional float32 array (including end-effector position, quaternion, gripper position, and base position). The generation pipeline involves Cosmos AR video to LeRobot v3.0 construction and droid IDM processing to output actions and states. It is designed for joint training with real datasets like Keh0t0/robocasa_mg30_real_v2 to enhance robotic learning.