SeonghuJeon/robocasa-v02-generated300-success-env-depth-rgb-256

RoboCasa v0.2 Generated-300 Success — Env Depth + RGB (256)数据集是基于RoboCasa v0.2 mg_im Generated-300 HDF5演示生成的，仅包含任务成功的演示。它覆盖了24个RoboCasa厨房原子任务，每个任务有300个经过成功审核的演示，总计约7,200个演示。数据集以256像素分辨率渲染环境深度和RGB图像，存储为.npy内存映射文件，旨在支持高吞吐量训练I/O。数据布局包括深度和RGB文件夹，每个文件夹包含任务、演示和摄像头的文件结构。摄像头配置包括robot0_agentview_left、robot0_agentview_right和robot0_eye_in_hand，分辨率为256×256，RGB和深度使用相同的相机内参和姿态。数据来源于上游HDF5发布，并通过成功审核过滤和渲染器生成。预期用途是为3DA / Shallow12 AR RoboCasa训练配置提供几何监督辅助数据，可与本地RoboCasa v3迷你仓库或v0.2 mg_im HDF5数据集配对使用。The RoboCasa v0.2 Generated-300 Success — Env Depth + RGB (256) dataset consists of per-step environment depth and RGB sidecars rendered from the original RoboCasa v0.2 mg_im Generated-300 HDF5 demos, restricted to demos whose final state matches the task-success predicate. It covers 300 success-audited demos per task across 24 RoboCasa Kitchen atomic tasks, totaling approximately 7,200 demos, rendered at native 256-pixel resolution. Geometry and RGB are emitted as per-camera contiguous .npy memmaps designed for high-throughput training I/O. The layout includes env_depth and env_rgb_256 directories with task/demo/camera file structures. Cameras include robot0_agentview_left, robot0_agentview_right, and robot0_eye_in_hand at 256×256 resolution, with identical intrinsics and poses per frame for RGB and depth. The source is the upstream RoboCasa v0.2 mg_im Generated-3000 release, filtered by per-task success audit and rendered via the 3DA exporter. Intended use is as drop-in geometric supervision sidecars for 3DA / Shallow12 AR RoboCasa training configs, paired with local RoboCasa v3 mini-repos or the v0.2 mg_im HDF5 dataset.