Tingxiang/robomme-tracking-masks

RoboMME Tracking Masks

Recommended training release: releases/native_scene_v1/train/.

新版训练标签：16 个任务、1600 段、768,897 帧。原待修的 40,670 帧已全部通过既定对齐协议并合入完整 episode；另对 VideoPlaceButton ep067 的 966 帧进行了整段实例身份修正。全量独立验收 1600/1600 通过。

This release contains per-object front-camera segmentation and tracking annotations paired with the original RoboMME training images. It preserves original frame order, actions, and annotation slots. Masks come from native simulator segmentation; they are not SAM predictions. Original RGB videos and robot actions are not included. Example review videos are included.

Versions

Location	Contents	Status
releases/native_scene_v1/train/<Task>/epNNN.npz	Complete repaired train episodes	Recommended; 1600 episodes, 768897 frames
Root train/	Historical v3 training labels	Preserved; known replay misalignment, superseded by the release above
Root test/	Historical test demonstrations	Unchanged; not part of this training repair or its validation

The original schema, logs, checksums and test notes still refer to the historical root files. See legacy README. New checksums and validation are release-local.

Download and read

from huggingface_hub import hf_hub_download, snapshot_download
import numpy as np

repo = "Tingxiang/robomme-tracking-masks"
p = hf_hub_download(repo, repo_type="dataset",
    filename="releases/native_scene_v1/train/VideoRepick/ep039.npz")
with np.load(p, allow_pickle=False) as z:
    t = 107
    masks = np.unpackbits(z["front_mask_bits"][t], axis=-1,
        bitorder="little").reshape(len(z["object_names"]), 256, 256).astype(bool)
    print(z["object_names"], masks.shape, z["accepted_frame"][t])

# Download only the new release (not the legacy train/test files):
snapshot_download(repo, repo_type="dataset",
    allow_patterns=["releases/native_scene_v1/**"], local_dir="robomme_masks")

Annotation format

Each NPZ is one complete episode. T is the frame count and N is the number of annotation slots. Object indices are local to each episode.

Array	Shape / interpretation
front_mask_bits	uint8 [T,N,8192], little-endian packed bits, 256×256 masks
object_names, object_kinds	[N], current logical object names and actor/link kinds
source_object_names	[N], historical slot names for traceability
native_instance_ids	[T,N], actual per-frame simulator IDs; use these instead of historical object_instance_ids
frame_index, frame_phase	[T]; original frame order, phase 0=demo / 1=execution
accepted_frame, frame_quality	[T]; all accepted, code definitions in quality_codes_json
pixel_area, visible	[T,N]
bbox_xyxy, uv_mask_center	[T,N,4], [T,N,2]
pose_world_xyz, pose_world_quat	[T,N,3] meters, [T,N,4]
front_intrinsic, front_extrinsic	Per-frame front camera calibration

The original task-relevant objects and articulation parts are retained, including dynamic highlights mapped to their logical owners. These are visible-surface masks: a fully occluded or not-yet-created object can have an empty mask. Background/table/floor are excluded. Invisible masks have zero area, NaN centroids and [-1,-1,-1,-1] boxes.

NPZ bytes are identical to the independently audited local release. Some embedded provenance strings retain historical collector filesystem paths; they are lineage information, not portable file dependencies. Use the release-relative manifest and hashes to locate the downloadable data.

Repair and validation

Missing hidden demo resets, differing initial layouts, object/highlight lifecycles, robot occlusion, camera pose precision and object identity swaps were addressed by restoring the simulated scene and extracting segmentation from the same observation. Frame-zero checks and method-specific normal controls preceded release qualification. Rejected-frame repairs additionally passed independent native-pixel, source alignment and negative-input checks.

All 40670 formerly rejected training frames were released under native_scene_rgb_deadband2_v1; the final full training set combines these with the previously accepted native labels. The whole 1600-episode assembly was independently checked for mask/pose lineage, frame order, slot names, visibility, areas, bounding boxes, centroids and current file hashes. VideoPlaceButton ep067 also received a verified full-episode identity override, already merged into the downloadable NPZ.

Acceptance means passing the documented alignment protocol, not byte-identical RGB or proof of a unique recovered latent state. Global RGB checks subtract a 2/255 per-channel quantization deadband before applying the 0.05 residual-mean threshold. Object-region RGB/depth and camera thresholds were not relaxed. Original images/actions are unchanged.

• Manifest and SHA256 per episode
• Final completion validation
• File checksums
• Failed-case comparison overview

Previous failed cases

Four columns: original RGB → original v3 → previous replay labels → final native masks. Centers were selected from previously rejected frames with large mask changes, not a random sample.

• VideoRepick ep039 video / image
• MoveCube ep084 video / image
• BinFill ep075 video / image
• ButtonUnmaskSwap ep035 video / image
• VideoPlaceButton ep067 video / image

Training coverage

Task	Episodes	Frames
BinFill	100	60,282
ButtonUnmask	100	26,626
ButtonUnmaskSwap	100	39,940
InsertPeg	100	47,703
MoveCube	100	39,388
PatternLock	100	20,782
PickHighlight	100	34,560
PickXtimes	100	53,720
RouteStick	100	36,900
StopCube	100	31,614
SwingXtimes	100	43,419
VideoPlaceButton	100	96,510
VideoPlaceOrder	100	112,392
VideoRepick	100	68,583
VideoUnmask	100	21,668
VideoUnmaskSwap	100	34,810

License

No new license is assigned by this upload. Refer to the original data provider’s terms for redistribution and reuse.

Test-demo annotations (labels only)

releases/test_demo_native_v1: 450 episodes, 182906 frames, approximately 77.4 MB of NPZ annotations with the legacy test schema. No RGB/depth bundle is included. Native consistency with regenerated demo images is checked; alignment to existing external test RGB remains unverified. Existing training release is unchanged.

Object and robot motion labels (five-frame window)

Latest motion annotations: releases/motion_window5_v1. Includes all 1,600 training episodes and 450 test demos: original object masks plus per-link robot masks, five-image-window motion states, and separate dynamic-object/dynamic-robot unions. Annotation-only; no RGB or depth payload. Older releases remain available. Test annotations refer to regenerated demos, not independently verified external test RGB.