armin1001/robocasa_dataset

RoboCasa Transfer Experiment Inputs

Input snapshot dated 2026-09-14 for the OpenCabinet, CloseCabinet and PickPlaceCounterToCabinet transfer experiments, including cached VLA features, demonstration data and simulator assets. This public repository contains the available experiment inputs, including the partial reference-task data described below. It is not a complete distribution of RoboCasa.

Model weights are supplied separately in the private armin1001/premodelrobocasa repository at revision 36f8ad7df1ffb59b6846ea90230d686dc9c1ced7. Access to that repository is required to restore the complete model and data input layout used by the experiment runners. See DOWNLOAD.md for pinned downloads, checksum verification and extraction instructions.

Contents

The available data snapshot contains 136,278 files / 27,338,576,924 bytes (about 25.5 GiB before compression), excluding the separately hosted models.

Archives restore these paths relative to a fresh AE_INPUTS directory:

Path	Contents
raw/<Task>/<date>/lerobot/	Available episode Parquets, camera videos and task metadata
queries/	Cabinet query metadata and the PickPlace demos15 demonstrations
features/	Existing rcoOC, rcoCC and rcorac feature caches and prior Gram data
robocasa_assets/	Kitchen, object and other RoboCasa simulator assets
robosuite_assets/	Robot and gripper simulator assets
robosuite_models_assets/	Additional simulator model assets
selection/	Historical input selection and its audit metadata
MANIFEST.md, README_pickplace_experiments.md	Original input provenance and experiment notes

The input_docs archive preserves the input documentation, including the historical manifest stored at raw/MANIFEST.md/20250819/lerobot. That unusual path is an inherited documentation artifact, not an additional task.

model/ and pretrained/ are excluded from this public dataset. Download them from the separate model repository above. Live experiment output checkpoints, evaluation videos, virtual environments and the executable code bundle are not included here.

Completeness and experiment requirements

The PickPlace target contains 108 demonstrations / 24,225 frames. Its seven source tasks contain 749 episodes / 210,185 frames. The demos15 query set contains 15 demonstrations / 3,347 frames. These are distinct roles: there is no queries/d108 in this snapshot.

The existing PickPlace, Target-PCA and DemInf experiments use the available inputs. DemInf-GP preparation remains incomplete: the reference prior needs raw Parquets and two camera streams for eight tasks that currently have only metadata and cached VLA features. The missing portion is 3,789 files for 1,263 episodes:

Task	Date	Missing files
PickPlaceStoveToCounter	20250819	327
PickPlaceMicrowaveToCounter	20250819	336
PickPlaceCounterToSink	20250819	324
PickPlaceCounterToOven	20250819	333
TurnOnStove	20250819	312
TurnOffStove	20250819	327
TurnOnSinkFaucet	20250819	321
NavigateKitchen	20250821	1,509

The snapshot-specific input preflight records ready: false and the missing paths. Uploading this snapshot does not supply those absent files. The input inventory records the actual available inputs; the archive manifest describes the packaged files.

Keep all feature caches. The original warm Oracle encoder checkpoint needed to recompute the existing VLA embeddings is not part of the base pre-model. The Cabinet query folders d15_OC and d15_CC contain metadata only; their selection uses these caches. demos15 includes raw data for re-embedding with DemInf. Available camera streams are preserved: the existing SIFT experiments use three cameras; DemInf and DemInf-GP use the hand camera and right agent-view camera.

Archive format

Each top-level input directory, or raw task, is a Zstandard-compressed Tar stream split into parts of at most 10 GiB. Names end in .tar.zst.part-0000, .tar.zst.part-0001, and so on. Parts are consecutive byte segments of one archive, not independently readable dataset shards. Concatenate all parts of an archive in numeric order before extracting.

SHA256SUMS provides transfer checksums. The archive paths are relative to AE_INPUTS; dereferenced source links are packaged as their contents so the restored data does not depend on links to the original cluster.

Provenance and notices

The original MANIFEST.md and README_pickplace_experiments.md describe the PickPlace target, pool, prior, query selection and cached feature extraction. Their historical counts and source descriptions are retained for provenance; use the snapshot inventory and preflight for present file availability. Per-task LeRobot meta/ files and query subset provenance are preserved.

The asset README files in robocasa_assets/ identify NVIDIA PhysicalAI Robotics Manipulation Objects Kitchen MJCF as their source. The included robosuite_models_assets/robots/aloha/LICENSE retains the Trossen Robotics notice, and robosuite_models_assets/robots/h1/LICENSE retains the Unitree Robotics BSD 3-Clause notice. Existing notices remain inside their original asset trees. This collection does not assign one new license to all underlying data, assets or models.

Running on another cluster

Use the matching experiment code snapshot and its README_cluster_migration.md to create the Python environments, configure the new cluster and launch jobs. After restoring both repositories, generate fresh prepared plans and datasets for that cluster. Historical selection metadata and source notes may contain old paths; they are provenance, not a portable executable job state. Preserve the runner's input and code hash checks, and rerun the DemInf-GP preflight before scheduling that experiment.