Datasets:

ChangChrisLiu
/

GWAM_Data

@@ -1,293 +0,0 @@
-# GWAM_Data v1.2 sparse-v2 schema
-This file is normative for the current GWAM_Data v1.2 sparse-v2 export.
-Old v1/v1.1 notes mentioning dense `rel_state`, dense visual tensors, `distractor` family index 1, or 317/323-D `node_feats` are historical only and must not be used to build v1.2 loaders.
-## Episode layout
-```text
-<split>/<task>/episode_000123/
-  _PHASE1_DONE.json
-  graph/
-    frames.jsonl
-    graph_static.json
-    node_state.npz
-    view_evidence.npz
-    dynamic_edges.jsonl
-    visible_masks_rle.jsonl.gz
-    visual_features_sparse.npz          # Phase 2 sparse visible-only features
-  audit/
-    summary.json                        # Phase 1 audit summary
-    phase2_summary.json                 # Phase 2 audit summary
-```
-Per-episode ZIPs should preserve the `<split>/<task>/episode_xxxxxx/` prefix.
-## Phase 1 tensors
-```text
-graph/node_state.npz
-  node_state: float32[T, 256, 32]
-  node_state_schema: str[32]
-```
-The `node_state_schema` array inside the npz is authoritative. Padded rows are all-zero; loaders must gate on the active channel before using any other field.
-Current v2 family ids:
-```text
-0 object, including distractors
-1 reserved / forbidden old distractor id
-2 surface
-3 receptacle
-4 fixture
-5 articulated_part
-6 handle
-7 control, defined but unpopulated in current export
-8 region, defined but unpopulated in current export
-9 robot
-```
-There is no target-relevance family bit in v2.
-## Static graph
-```text
-graph/graph_static.json
-```
-Contains deterministic node inventory, schema/rule-version stamps, source metadata, and supervision-only goal metadata. The goal/instruction/object-role block is not a model-input tensor.
-Required schema stamp:
-```text
-GWAM_Data_v1.2_phase1_sparse_v2
-```
-All packaged episodes should share one identical `rule_versions` blob.
-## View evidence and masks
-```text
-graph/view_evidence.npz
-graph/visible_masks_rle.jsonl.gz
-```
-Mask contract:
-```text
-views: side_right, side_left, wrist
-side_right -> robot0_agentview_right
-side_left  -> robot0_agentview_left
-wrist      -> robot0_eye_in_hand
-min_visible_px: 20
-RLE: row-major binary RLE v2_vectorized
-resolution: 256x256
-```
-RLE masks and `view_evidence` visible pairs must agree exactly.
-## Dynamic edges
-```text
-graph/dynamic_edges.jsonl
-```
-Current sparse-v2 event vocabulary:
-```text
-contact
-visibility_change
-```
-There is no dense `rel_state[T,N,N,C]` tensor in v1.2 sparse-v2.
-## Phase 2 sparse visual features
-Phase 2 uses the Phase 1 visible masks and writes sparse features only for visible node/view pairs.
-Required invariants:
-```text
-no invisible features
-valid_pairs union invalid_pairs == Phase 1 visible pairs
-valid_pairs intersect invalid_pairs == empty
-no duplicate pairs
-valid feature vectors are fp16 and unit-normalized
-type text vectors are deterministic across episodes
-```
-Phase 2 writes:
-```text
-graph/visual_features_sparse.npz
-  t int32[P], n int16[P], v int8[P]
-      valid sparse pair indices: frame, node, view
-  feat float16[P, 256]
-      unit-normalized visible-only visual features
-  type_clip32 float32[256, 32]
-      per-node-slot projected CLIP text type vectors
-  invalid_t, invalid_n, invalid_v
-      visible Phase-1 pairs where feature extraction failed
-  visual_computed bool[T]
-      per-frame Phase-2 coverage flag
-  cameras str[3]
-      view names
-  n_max int32[1]
-      fixed node cap, 256
-audit/phase2_summary.json
-  Phase 2 audit counts and visible-pair partition summary.
-```
-Do not expect dense `[T,N,V,D]` tensors.
-## Model-input discipline
-Allowed default model input:
-```text
-node_state active rows
-sparse dynamic_edges
-Phase 2 visible-only features and their pair indices
-robot/action streams if explicitly included by experiment
-```
-Supervision/metadata only unless explicitly allowed:
-```text
-instruction
-fixture_refs
-object_roles
-goal metadata
-node_id strings
-source strings
-frames.jsonl timestamp/index/next.done alignment fields
-```
-Do not use raw node slot position as a semantic cue; node ordering can correlate with task-object ordering.
-## Live real-time environment graph compatibility
-The on-disk release stores full episodes, while the real-time extractor returns one current simulator frame at a time. They use the same schema vocabulary:
-```text
-on-disk episode:        node_state float32[T,256,32]
-live current frame:     node_state float32[256,32]
-live K-frame buffer:    node_state float32[K,256,32]
-```
-The live extractor in `realtime/gwam_realtime_env_graph.py` / `extractors/gwam_realtime_env_graph.py` uses the same:
-```text
-node_state_schema
-family ids
-active-mask policy
-static prior edge labels: part_of, controls
-dynamic edge labels: contact, visibility_change
-view_evidence fields and normalized [0,1] units
-row-major RLE mask codec when masks are requested
-```
-Real-time extraction is for users with their own RoboCasa/robosuite environment and assets. It constructs a current-observed graph from the current simulator state and segmentation renders. It does not compute the offline Phase-2 SAM2/CLIP `visual_features_sparse.npz` by default. Future simulator states, future robot nodes, future contacts, goal-derived target-node ids, object role labels, and instruction-derived target edges are not default-safe evaluation inputs; use them only in explicitly labeled oracle or teacher-forced modes.
-## Known limitations
-1. `target/OpenDrawer/episode_000459` is dropped because source env initialization failed repeatedly.
-2. Six target tasks are state-blind in `node_state`: CloseFridge, OpenStandMixerHead, SlideDishwasherRack, TurnOffStove, TurnOnElectricKettle, TurnOnMicrowave. Use a future `fixture_state.npz` sidecar or document these as geometric/contact/visibility-supervised only.
-3. `control` and `region` family ids are defined but unpopulated in the current export.
-4. `N_MAX=256`; max observed node count is close to the cap.
-## Normative codec, unit, and feature provenance addendum
-### Frame/action alignment
-`frames.jsonl` maps graph timestep `t` to source `source_frame` 1:1 with no subsampling. RoboCasa LeRobot metadata records 20 Hz (`record_frequency=20.0`), so frame timestamps advance by 0.05 s. For the PandaOmron source episodes, `meta/modality.json` describes the 12-D action layout as `base_motion[0:4]`, `control_mode[4:5]`, `end_effector_position[5:8]`, `end_effector_rotation[8:11]`, and `gripper_close[11:12]`.
-### `view_evidence.npz` units
-All image-space values are normalized to [0,1] in the 256x256 upright image frame: `view_centroid=(mean_col/256, mean_row/256)`, `view_area=pixels/(256*256)`, and `view_bbox=[x_min,y_min,x_max,y_max]/256` from inclusive pixel bounds. Camera intrinsics/extrinsics are not exported; both agentview cameras move with the mobile base and the wrist camera moves with the hand.
-### `visible_masks_rle.jsonl.gz` codec
-Rows are gzipped JSONL records of the form `{"t": int, "masks": [{"n": int, "v": int, "area_px": int, "bbox": [x_min,y_min,x_max,y_max], "rle": [runs...]}]}`. Frames with zero visible pairs are omitted. The mask is 256x256, flattened row-major from the upright image, and `rle` is alternating run lengths starting with the background/zero run; runs sum to 65,536.
-### Phase-2 visual features
-`feat` is a 256-D float16 SAM2.1 `hiera_base_plus` image-encoder feature pooled under the Phase-1 visible RLE mask on the source LeRobot RGB frame, then L2-normalized. `invalid_t/n/v` lists Phase-1 visible pairs that had no support after downsampling to the SAM2 embedding grid; no feature is fabricated. `type_clip32` is a 32-D CLIP ViT-B/32 category-text feature generated with prompt `a photo of a {category_text} in a kitchen.`, projected by seeded Gaussian random projection seed 20260702 and L2-normalized. Padded node slots are not zero in `type_clip32`; always gate on `node_state[..., active]` before consuming it.
-Note: padded node slots are not zero in `type_clip32`; active gating is mandatory for `type_clip32` as well as `node_state`.
-## Structured model representation addendum
-Node tensor:
-```text
-node_state: float32[T,256,32]
-active_mask: bool[T,256] = node_state[...,0] == 1
-```
-Node-state typed channel groups:
-```text
-0 active: binary mask
-1 family_idx: categorical id 0..9
-2:4 static flags: free_movable, articulated
-4:11 pose: xyz + quat_xyzw
-11:17 velocity: linear xyz + angular xyz
-17:19 scalar joint state: qpos, qvel
-19:22 visibility flags: right, left, wrist
-22:32 reserved zeros
-```
-Static/prior edge tensors derived from `graph_static.json/prior_edges`:
-```text
-prior_edge_index: int64[2,E_prior]
-prior_edge_type: int64[E_prior] where part_of=0, controls=1
-prior_edge_attr_onehot: float32[E_prior,2] optional
-```
-Dynamic/event edge tensors derived per JSONL row from `dynamic_edges.jsonl`:
-```text
-dyn_edge_index[t]: int64[2,E_dyn_t]
-dyn_edge_type[t]: int64[E_dyn_t] where contact=0, visibility_change=1
-dyn_edge_attr[t]: optional concat(rel_onehot[2], view_onehot[3], is_self_edge[1])
-```
-`contact` edges are stored as sorted node pairs and are physically undirected; add reverse edges explicitly for bidirectional message passing. `visibility_change` edges are self-edges with a required `view` id. No force/normal/contact-point feature is stored.
-Static node records (`node_id`, `category_text`, `root_bodies`, `source`) are metadata; do not use string ids/source paths as default model inputs. `type_clip32[256,32]` is an optional category-text node feature and must be active-gated because padded slots are non-zero.
-## Online final-graph extraction files
-The public real-time inference/evaluation path can construct the full final graph for one current simulator step:
-```text
-extract_final_graph(visual_backend=Sam2ClipRealtimeFeatureBackend(...))
-current env state/RGB/segmentation/masks + Phase 1 graph + Phase 2 SAM2/CLIP = x [N_real,342]
-```
-Bundled direct-run files:
-```text
-realtime/gwam_realtime_env_graph.py
-examples/realtime_env_graph_eval_loop.py
-scripts/setup_realtime_final_graph_env.sh
-scripts/verify_realtime_final_graph.py
-models/sam2/checkpoints/sam2.1_hiera_base_plus.pt
-models/clip/ViT-B-32.pt
-models/MODEL_MANIFEST.json
-```
-The default online final graph feature layout is:
-```text
-20 state dims + 10 family one-hot + 256 SAM2 masked pooled visual + 32 type_clip32 + 24 view evidence = 342
-```
-`extract_current_graph()` is a lightweight Phase-1/debug path only and is not the final model-input graph.