SeonghuJeon/robocasa-v02-generated300-success-env-depth-rgb-256
RoboCasa v0.2 Generated-300 Success — Env Depth + RGB (256) 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. The selection covers 300 success-audited demos per task across 24 RoboCasa Kitchen atomic tasks (~7,200 demos total) at native 256-pixel render resolution. Geometry and RGB are emitted as per-camera contiguous .npy memmaps designed… See the full description on the dataset page: https://huggingface.co/datasets/SeonghuJeon/robocasa-v02-generated300-success-env-depth-rgb-256.
RoboCasa v0.2 Generated-300 Success — Env Depth + RGB (256)
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. The selection covers 300 success-audited demos per task across 24 RoboCasa Kitchen atomic tasks (~7,200 demos total) at native 256-pixel render resolution. Geometry and RGB are emitted as per-camera contiguous .npy memmaps designed for high-throughput training I/O.
Layout
env_depth/
manifest.json # merged canonical manifest for the depth set
env_depth_memmap/<task>/<demo>/<camera>.depth.npy
env_rgb_256/
env_rgb_memmap/<task>/<demo>/<camera>.rgb.npymanifest.json enumerates each task / demo / camera triple with frame counts, success flags, and the source HDF5 path/demo key. The depth sidecars carry exact MuJoCo z-buffer geometry rendered with site markers hidden, matching the rollout protocol used by 3DA training.
Cameras
robot0_agentview_left
robot0_agentview_right
robot0_eye_in_hand256×256 resolution. RGB and depth use identical camera intrinsics and poses per frame.
Source
- Upstream HDF5: RoboCasa v0.2
mg_imGenerated-3000 release (robocasa_v0.2_generated3000_mg_im). - Demo filter: per-task success audit on the final environment state, prioritized by the upstream
mask/300_demosgroup, then 450 additional candidates per task, trimmed to 300 successful demos per task. - Renderer: RoboCasa / robosuite / MuJoCo via the 3DA exporter (
scripts/export_robocasa_v02_generated300_depth_*), site markers hidden, identical camera config to the policy rollout adapter.
Intended use
Drop-in geometric supervision sidecars for 3DA / Shallow12 AR RoboCasa training configs, paired with the local RoboCasa v3 mini-repos or the v0.2 mg_im HDF5 dataset. Not affiliated with NVIDIA's official Cosmos Policy release.
