ajaysri/sandwich_v1_human_v1

Sandwich v1 human dataset

This public review release contains 4,000 transfer episodes from 1,000 generated hand-proxy sandwich recordings. Each source recording has four ordered ingredient transfers. data/episodes.parquet has one row per transfer, with its active object, source action/state/image interval and separate 5 fps prep-camera clip. This matches the robot dataset's one transfer per exported episode boundary. The 1,000 complete source recordings are preserved separately in data/sources.parquet, full-length preview videos and lossless archives.

The hand is a scripted kinematic visual proxy in simulation, not a physical human or a human-operated robot. There are no robot actions in this dataset. This is a public review export, not yet a labeled HLP training table or evidence of model performance. Training awaits owner approval and prelaunch audits.

Browse one source by transfer

Transfer	Active object	Prep-camera clip
0	bread	episode 0000 transfer 0
1	lettuce	episode 0000 transfer 1
2	tomato	episode 0000 transfer 2
3	onion	episode 0000 transfer 3

See also the whole recording and its goal card. The four clip links above are review previews. The source's exact prep JPEGs and all saved state data are in its lossless archive shard.

Index and timing

Each transfer spans 112 control actions at 20 Hz. The saved prep frames are every fourth state, so each transfer clip has 28 frames at 5 fps. Source action and observation-state intervals are half-open: [0,112), [112,224), [224,336), [336,448). The final state at source index 448 has no following action; it remains in the complete source recording and full-length preview but is excluded from transfer clips. No action is dropped or duplicated.

data/episodes.parquet columns include episode_index (0–3999), source_episode, subtask_index (0–3), active_object, source action/state/prep-frame ranges, condition labels, clip and goal paths, and archive path. data/sources.parquet has one row per complete source recording. meta.json gives counts. The full source video is at videos/prep/episode_XXXX.mp4; transfer clips are at videos/prep/transfers/episode_XXXX_transfer_K.mp4.

Each archives/shard_XX.tar contains original raw/episode_XXXX/ and derived prep/episode_XXXX/ files. The raw folder includes the recording, per-action trajectory, saved simulator states, original goal-side JPEGs and goal card. The prep folder includes the approved prep-camera JPEGs and a derivation record. The archives are lossless provenance; the H.264 MP4s are convenient previews. Sources are assigned to archive shard by source episode number modulo 16.

Conditions and checks

There are 200 robot-familiar source recordings and 200 source recordings each for bread-region, ingredient-combination, stack and avocado conditions. Every source has bread plus three toppings and follows direct placement. Avocado is human-only training exposure; it is absent from robot training data. Four transfer episodes inherit their source condition label, so counts of source recordings and counts of transfer episodes must not be confused.

All 1,000 source recordings passed the approved sandwich_splits_v1 recording-admission predicate. Every source trajectory was checked to contain four contiguous 112-action active-object runs, beginning with bread, with matching saved states and 5 fps frame steps. The approved prep-camera rerender binds its images to the source recording and state hashes; the export rechecked these links. The 16 source archives passed SHA-256 validation. The transfer clips and index were checked for complete coverage, exact frame counts and nonoverlapping ranges.

The generated hand can move outside the prep-camera frame; this known visibility limitation needs owner review before HLP target construction. No HLP target rows, train/validation split, complete checkpoint-lineage audit or learned-policy results are included. Simulator state may inform training targets, but a deployed command provider must use permitted observations rather than privileged state.