ajaysri/sandwich_v1_steerable_llp_v1

Sandwich v1 steerable low-level policy dataset

Public review release. Training awaits owner approval and a final independent audit. This dataset contains 100 successful human-operated simulated UR5e/WSG50 sandwich demonstrations. Each source demonstration has four ingredient transfers. Each transfer is exported once for each command interface: language, point, and drawn tool path. Thus there are 1,200 LeRobot v3 episodes, 400 per interface, and 322,848 rows at 20 Hz. The three interface copies share the same state and action trajectory. There is no evaluation split or learned-policy result.

What is an episode?

One exported episode is one ingredient transfer, including its current prep/wrist frames, action sequence, and periodically refreshed command. A source demonstration becomes 12 episodes (4 transfers × 3 interfaces). No source action frames were dropped: each interface has 107,616 action rows across all 100 sources. The final retreat is retained in transfer 4. meta/conditioning.parquet gives each episode's source demonstration, interface kind, subtask number (0 to 3), and source frame interval. The observation.conditioning.* row fields provide corresponding provenance and are not policy inputs.

Frame fields

Field	Meaning
observation.images.main	Current prep-camera image, 240 × 160 RGB video.
observation.images.wrist	Current wrist-camera image, 240 × 160 RGB video.
observation.images.goal	Prep-camera command image held from the latest command anchor.
observation.state	Six robot joints and jaw opening.
action	Six joint targets and gripper opening control.
prompt	Language command or visual-interface instruction.
observation.conditioning.source_frame	Frame index in the original four-transfer source demonstration.
observation.conditioning.anchor_frame	Command anchor index within this transfer episode.

Commands refresh every 20 control steps and when the transfer changes. The last command remains held during the final retreat. frame_index and anchor_frame are local to the exported transfer episode.

• Language: deterministic ingredient-placement sentence. The goal video stores an unmarked anchor image for inspection. Training must mask the goal-image slot on language rows; that image is not part of the language interface.
• Point: blue marks the active object and pink its destination. The blue dot remains on the object when grasped, since the command image may remain stale while the low-level policy acts. The anchor image is held between updates.
• Path: a future tool-centre trace, with gripper open/close tags, clipped to the current transfer.

Prep, wrist, and goal streams are H.264 video. States, actions, prompts, and provenance are in Parquet. meta/stats.json has exact corpus statistics for state and action, including quantiles. Use a LeRobot v3-compatible reader (validated with LeRobot 0.4.4).

Scope and review

Scenes contain bread and three toppings, with no avocado in robot training. Collection follows direct placement: flat starts, bread first, bottom-up assembly, one final transfer per object, with no parking or destacking. Source recordings passed robot-training admission under the approved sandwich_splits_v1 definition. These trajectories are training exposure for later evaluation-lineage checks.

The release was checked for 1,200 complete transfer episodes, all 100 sources × 4 transfers × 3 interfaces, contiguous source-frame coverage, local anchors inside their transfer, equal interface balance, and exact state/action quantiles. All 322,848 state/action rows match the previous whole-task export exactly after aligning by source, source frame, and interface. The visual streams are separately encoded H.264 and need visual inspection. A final independent audit remains pending before training.

Simulator state helped construct targets and commands; a deployed command provider may not use it. At evaluation time, commands must come from a human, a learned high-level policy, or a labeled oracle diagnostic. This dataset does not establish closed-loop success or out-of-distribution performance.