robocasa-H50-256

The RoboCasa human-50 training demos of DAVIAN-Robotics/robocasa-H50 (24 atomic kitchen tasks, 1,293 episodes, 357,335 frames), re-rendered at 256×256 with the same randomized cameras and the same generative textures as RoboCasa's official demo_gentex_im128_randcams.hdf5 training release. Only the image resolution differs from H50.

What is identical to robocasa-H50

Episode order, action (12-D), observation.state (16-D), next.done, timestamp, frame_index, episode_index, index, task_index and the task table are copied from H50 and are bit-identical to it. Video codec (av1, yuv420p) and fps (20) are the same; video resolution is 256×256 instead of 128×128.

How the images were made

RoboCasa's official gentex/randcams render stores each episode's randomized camera poses (ep_meta.cam_configs) but not its generative-texture draw (gen_textures is empty and the stored model XML keeps default textures; the draw came from an unseeded RNG). The draw is one index in [0, 100) selecting a (cabinet, counter, floor, wall) texture set, so for every episode:

1. Texture recovery: frame 0 was rendered at 128 with the official cameras for candidate sets and matched against the official 128 frame (the correct set is an outlier: MAE ≈ 1 vs ≥ ~20 for every other set).
2. Gate: with the recovered textures and official cameras, the first / middle / last frame were re-rendered at 128 on GPU and compared with the official 128 frames. All 1,293 episodes passed: worst case 0.2% of pixels off by more than 2 intensity levels (isolated z-fighting pixels); most frames are bit-identical.
3. Render: every frame was rendered at 256 with the same states, cameras and textures.

Replayed from the recorded simulator states (demo.hdf5 states), so object and robot configurations are exactly the official ones.

Format

LeRobot v3.0, robot_type robomimic, 1293 episodes, 357335 frames, 242 tasks, cameras observation.images.robot0_{agentview_left,agentview_right,eye_in_hand}_image.