docs/EXTRACTION_PLAN.md

# Extraction Plan: RoboCasa target Atomic-Seen Graph Dataset v1

## Inputs

RoboCasa target Atomic-Seen human datasets, 18 tasks:

- `CloseBlenderLid`: `v1.0/target/atomic/CloseBlenderLid/20250822`, horizon 900
- `CloseFridge`: `v1.0/target/atomic/CloseFridge/20250816`, horizon 900
- `CloseToasterOvenDoor`: `v1.0/target/atomic/CloseToasterOvenDoor/20250818`, horizon 450
- `CoffeeSetupMug`: `v1.0/target/atomic/CoffeeSetupMug/20250813`, horizon 600
- `NavigateKitchen`: `v1.0/target/atomic/NavigateKitchen/20250821`, horizon 450
- `OpenCabinet`: `v1.0/target/atomic/OpenCabinet/20250813`, horizon 1050
- `OpenDrawer`: `v1.0/target/atomic/OpenDrawer/20250816`, horizon 750
- `OpenStandMixerHead`: `v1.0/target/atomic/OpenStandMixerHead/20250818`, horizon 450
- `PickPlaceCounterToCabinet`: `v1.0/target/atomic/PickPlaceCounterToCabinet/20250811`, horizon 750
- `PickPlaceCounterToStove`: `v1.0/target/atomic/PickPlaceCounterToStove/20250818`, horizon 600
- `PickPlaceDrawerToCounter`: `v1.0/target/atomic/PickPlaceDrawerToCounter/20250820`, horizon 750
- `PickPlaceSinkToCounter`: `v1.0/target/atomic/PickPlaceSinkToCounter/20250813`, horizon 900
- `PickPlaceToasterToCounter`: `v1.0/target/atomic/PickPlaceToasterToCounter/20250817`, horizon 600
- `SlideDishwasherRack`: `v1.0/target/atomic/SlideDishwasherRack/20250820`, horizon 450
- `TurnOffStove`: `v1.0/target/atomic/TurnOffStove/20250812`, horizon 750
- `TurnOnElectricKettle`: `v1.0/target/atomic/TurnOnElectricKettle/20250817`, horizon 450
- `TurnOnMicrowave`: `v1.0/target/atomic/TurnOnMicrowave/20250813`, horizon 450
- `TurnOnSinkFaucet`: `v1.0/target/atomic/TurnOnSinkFaucet/20250812`, horizon 600

## V1 outputs

For each episode, export:

```text
RGB: side_right, side_left, wrist
Segmentation/masks: side_right, side_left, wrist
Robot state: T x 16
Action: T x 12
Reward/done/success
Simulator state: qpos/qvel, body poses, contacts
Graph: static nodes/prior edges + per-frame state edges + event labels + tensors
```

Depth is not exported in v1.

## Download command

RoboCasa downloader supports:

```bash
python -m robocasa.scripts.download_datasets   --split target   --source human   --tasks <task names>
```

For the 18 target Atomic-Seen tasks, pass the task list from `docs/TARGET_ATOMIC_SEEN_TASKS.md`.

Important: the downloader prompts for confirmation. Do not pipe automatic yes in shared/safety-managed environments. Confirm manually after checking disk.

Current local inspection found about 562G free under `/home/chris` on 2026-07-02, but actual download size was not measured because automatic confirmation to the downloader was blocked by the safety gate. Use a manual dry run or inspect Box links before full download.

## Replay/extraction loop

For each raw HDF5 episode:

1. Load `states`, `actions`, `model_file`, and `ep_meta`.
2. Create RoboCasa env from dataset metadata.
3. For each timestep `t`:
   - set simulator state to `states[t]`;
   - call `sim.forward()`;
   - render RGB for:
     - `robot0_agentview_right` -> `side_right`,
     - `robot0_agentview_left` -> `side_left`,
     - `robot0_eye_in_hand` -> `wrist`;
   - render MuJoCo segmentation for the same cameras;
   - map segmentation geom IDs to body IDs and graph node IDs;
   - read `sim.data.qpos`, `qvel`, `xpos`, `xquat`, contacts;
   - read/reconstruct robot state and action;
   - assemble per-frame graph record.
4. After episode, write graph tensors and event labels.
5. Run sanity checks before accepting episode.

## Graph inventory construction

Inventory is frozen per episode:

```text
object nodes: task objects + distractors
surface nodes: counters, sink surfaces, stove top, etc.
receptacle nodes: cabinets, drawers, fridge, sink basin, toaster oven, etc.
fixture nodes: appliances/fixtures involved in task
articulated_part nodes: doors, drawers, lids, racks, knobs/switches where relevant
handle/control nodes: handles, knobs, faucet handle, buttons where visible/controllable
optional robot node: off by default
```

Use deterministic `node_grouping_<task>.yaml`; do not let language decide which nodes exist.

## Contact filtering

Filter raw MuJoCo contacts by:

- geom-to-node mappability;
- dropping fixture-fixture static contacts;
- aggregating multiple geom contacts into one node pair;
- recording count/min distance;
- keeping robot-object contacts separately from object-surface support.

## Segmentation output

Each segmentation file should map node IDs to masks:

```text
seg/side_right/frame_000123.npz
  node_ids: string array
  masks_rle or masks_bool packed
  geom_ids_by_node
  visible flags
```

## Acceptance gates

An episode is accepted only if:

- RGB, segmentation, state, action, and graph frame counts match exactly.
- Every visible mask maps to a known node or documented ignored geom class.
- `rel_state` contains no prior or goal channels.
- `rel_prior` contains no instruction-specific target edges.
- State/action shapes are `(T, 16)` and `(T, 12)`.
- Required camera files exist for side_right, side_left, wrist.
- Per-task event expectations are roughly satisfied, or the episode is quarantined with reason.