| { |
| "schema": "vpm-worldbench/sample/1.0", |
| "uid": "Real-Robot/episode_034", |
| "subset": "Real-Robot", |
| "id": "episode_034", |
| "source": { |
| "dataset": "RB-Y1 WujiHand2 teleop ego+exo (100 episodes) + 3D point-track re-export", |
| "repo": "hf:kdh8156/RB-Y1_WujiHand2_teleop_ego_exo_100", |
| "revision": "269c3c008f0990f00ee86698c468a51a0117e959", |
| "tracks_repo": "hf:rooty2020/RB-Y1_WujiHand2_teleop_ego_exo_100_pointtracks", |
| "tracks_revision": "b3d7602562cd5cba2a11d37dbbbf299916796d84", |
| "id": "episode_034", |
| "episode_index": 34, |
| "license": "cc-by-nc-4.0 (pointtracks dataset card); upstream kdh8156 licence not recorded" |
| }, |
| "n_views": 2, |
| "n_frames": 202, |
| "fps": 15.0, |
| "duration_s": 13.4667, |
| "views": [ |
| { |
| "name": "view0", |
| "video": "videos/view0.mp4", |
| "width": 1280, |
| "height": 720, |
| "source_name": "exo", |
| "static": true, |
| "hfov_deg": 101.11, |
| "vfov_deg": 68.72, |
| "role": "exo", |
| "camera_model": "ZED 2i", |
| "serial": 32803500, |
| "lens": "left, rectified (no distortion)", |
| "mount": "external tripod facing the robot across the table", |
| "source_video": "exo.mp4", |
| "lerobot_video_key": "observation.images.camera_exo_left", |
| "calibration": "static T_exo_base (camera_calibration.json: ego->exo ZED relative extrinsic composed with the ego->base fit; README: exo projection exact); one calibration for all 100 episodes" |
| }, |
| { |
| "name": "view1", |
| "video": "videos/view1.mp4", |
| "width": 1280, |
| "height": 720, |
| "source_name": "ego", |
| "static": true, |
| "hfov_deg": 82.23, |
| "vfov_deg": 52.3, |
| "role": "ego", |
| "camera_model": "ZED-M", |
| "serial": 15142661, |
| "lens": "left, rectified (no distortion)", |
| "mount": "robot head, fixed pitch 45 deg (head_0 = 0, head_1 = 0.7854 rad); static in the base frame", |
| "source_video": "ego.mp4", |
| "lerobot_video_key": "observation.images.camera_ego_left", |
| "calibration": "static T_ego_base = the dataset's manual RGB fit to the robot (camera_calibration.json robot_frame_fit: IoU ~0.82 train, 0.77-0.81 held-out; README: small far-left residual). The nominal head-FK pose (labels/ego_fk_c2w.npy) differs by ~2.1 deg / ~6 cm and is not used for projection" |
| } |
| ], |
| "world": { |
| "description": "RB-Y1 robot base frame, unchanged: x forward (toward the table and the exo camera), y = robot's left, z up, metres, gravity -z. The mobile base never moves (planar twist and wheels 0), so it is one static frame shared by all 100 episodes; the table top is at z ~ 0.8 m (floor height not documented).", |
| "units": "m", |
| "handedness": "right", |
| "up": "+z", |
| "to_source": { |
| "note": "X_source = A @ X + b (spatial map; the time map is tracks.time_offset_frames)", |
| "A": [ |
| [ |
| 1.0, |
| 0.0, |
| 0.0 |
| ], |
| [ |
| 0.0, |
| 1.0, |
| 0.0 |
| ], |
| [ |
| 0.0, |
| 0.0, |
| 1.0 |
| ] |
| ], |
| "b": [ |
| 0.0, |
| 0.0, |
| 0.0 |
| ], |
| "source_units": "m", |
| "source_frame": "robot base (meta.json point_frame)" |
| } |
| }, |
| "files": { |
| "cameras": "cameras.npz", |
| "tracks3d": "tracks3d.npz", |
| "captions": "captions.json", |
| "depth": null, |
| "labels": { |
| "action": "labels/action.parquet", |
| "proprio_30hz": "labels/proprio_30hz.parquet", |
| "ego_fk_c2w": "labels/ego_fk_c2w.npy" |
| } |
| }, |
| "tracks": { |
| "n_points": 768, |
| "groups": [ |
| "arm", |
| "hand" |
| ], |
| "default_report_group": "all", |
| "time_offset_frames": -2.3, |
| "time_alignment": "xyz(t) = (1-a) * FK[floor(s)] + a * FK[floor(s)+1] with s = t - 2.3, a = s - floor(s); s < 0 clamps to FK frame 0 (robot static; motion onset at FK frame 3-4). The original FK tracks are source/robot.npz + source/hand.npz.", |
| "layout": "0..511 = source/robot.npz points[:, i] (arm: right 0..232, left 233..511); 512..767 = source/hand.npz points[:, i - 512] (hand: right 512..645, left 646..767)", |
| "parts": "part = MJCF body of the rigid arm link block (link_<side>_arm_k, k = link_block = 0..6 proximal -> distal, driven by joint <side>_arm_k) or hand_right / hand_left (finger membership not labelled)", |
| "dup_of": "dup_of[i] = j < i when point i duplicates point j (593 -> 591 exact; 301 -> 294 and 481 -> 470 within 2.4 um); -1 otherwise", |
| "visibility_rule": "in_frustum = valid & camera z > 0.05 m & projection inside the 1280x720 image (pixel centres at integers, u in [-0.5, 1279.5), v in [-0.5, 719.5)). visible = in_frustum & not self-occluded: point z-buffer at the native 1280x720 resolution with 12x12 px cells (grid aligned to the image edge at -0.5), per cell the minimum camera depth over all in-frustum points of the 768 (arm + hand), and a point is visible when its depth <= the minimum over its 3x3 cell neighbourhood + 0.03 m (the pre-release z-buffer visibility rule, 8 px cells at 480x848 -> 12 px native). Only robot self-occlusion by the sampled arm/hand surface points is modelled: scene objects (basket, pot, cups, ...) and unsampled robot parts (torso, head) never occlude. Computed on the time-aligned xyz (tracks.time_offset_frames).", |
| "accuracy": "Forward kinematics of the measured joint positions (LeRobot observation.joint_position) through the dataset MJCF (rby1m_wujihand2): exact up to the robot model and encoder error. Projection error is dominated by the camera calibration: exo is exact per the source README, except that exo frame 0 of episodes 000-022 sits ~1.8 px lower than in 023-099 with identical K/T (~2 px extrinsic error, flag calibration_step_exo); the ego extrinsic is a manual fit (IoU ~0.8, small far-left residual, flag ego_extrinsic_manual_fit). Time: the videos trail the FK by ~2.3 frames; one global offset of -2.3 frames is applied to both views (Real-Robot/_source/rby1_lag.json, 200 episode-views: median lag exo -2.50, ego -2.06; per-episode two-view mean -2.28 +- 0.19). The residual around -2.3 (measured lag + 2.3; < 0 = the video still trails the released tracks, > 0 = it leads them) is view-dependent: exo -0.21 and ego +0.25 frame on average (medians -0.20 / +0.24), with a per episode-view spread of std 0.33 (exo) / 0.26 (ego) frame and 5-95 % range exo -0.71..+0.26, ego -0.19..+0.68 (at peak_corr >= 0.8: exo -0.81..+0.38, ego -0.38..+0.68; the extremes exo -1.13, ego +1.00 have peak_corr < 0.8). Per-sample values: attributes.measured_lag_frames; flag time_offset_residual_gt_0.5:<view> where |residual| > 0.5 frame with peak_corr >= 0.8 (32 of 200 episode-views). A residual of r frames displaces a point by r x its FK step per frame (max 3.2 cm/frame, fastest hand motion): <= 0.65 cm at 0.2 frame, <= 1.6 cm at 0.5 frame. Linear interpolation between 15 Hz FK samples adds < 1 mm. Visibility is approximate (self-occlusion only). Points 591/593 are exact duplicates and 294/301, 470/481 coincide within 2.4 um (dup_of).", |
| "visible_outside_frustum_dropped": 0, |
| "extras": { |
| "dup_of": "(N,) int16, j < i when point i duplicates point j (593 -> 591 exact; 301 -> 294 and 481 -> 470 within 2.4 um); -1 otherwise", |
| "link_block": "(N,) int8 rigid arm link block 0..6 (proximal -> distal) of arm points; -1 for hand points" |
| } |
| }, |
| "depth": null, |
| "caption": "Put the basketball on the plate with the right hand.", |
| "attributes": { |
| "difficulty": "normal", |
| "category": "put_on_plate", |
| "hand": "right", |
| "task_family": "put_on_plate", |
| "task_variant": null, |
| "n_steps": 1, |
| "robot": "rby1m_wujihand2", |
| "episode_index": 34, |
| "fps_original": 30, |
| "orig_frame_index": "video/camera frame t = original 30 Hz LeRobot frame 2t (videos and FK were subsampled every 2nd frame; labels/proprio_30hz.parquet row 2t = labels/action.parquet row t = source FK frame t). Released xyz(t) = FK at 30 Hz frame 2*(t - 2.3), clamped at 0.", |
| "measured_lag_frames": { |
| "exo": { |
| "lag_frames": -2.51, |
| "peak_corr": 0.924, |
| "residual_after_offset": -0.21 |
| }, |
| "ego": { |
| "lag_frames": -2.18, |
| "peak_corr": 0.833, |
| "residual_after_offset": 0.12 |
| }, |
| "mean_lag_frames": -2.345, |
| "note": "lag L maximises corr(video frame-difference energy in the projected-robot mask, projected FK point speed at t+L) on the ORIGINAL (unaligned) FK tracks (Real-Robot/_source/rby1_lag.json); L < 0 = video trails the FK. Released tracks use one global offset -2.3." |
| }, |
| "n_default_queries_at_ref_frame": 432, |
| "labels_time_offset_frames": -2.3, |
| "labels": { |
| "action": "action.parquet: 15 Hz, row t = source FK frame t (= LeRobot row 2t); columns 'action' (66 absolute joint-position targets, rad, names in _source/info.json features.action.names), 'action.mobile_base.planar_twist' (3, all 0), 'action.pre_filter' (66), 'action.filter_applied' (66 bool). Restored from the HF pointtracks tar.", |
| "proprio_30hz": "proprio_30hz.parquet: the raw 30 Hz LeRobot episode parquet (joint position/velocity/effort, force/torque, wrist and camera poses in the base frame, next.success/done, timestamp); row 2t = source FK frame t; 403 rows (2T-1).", |
| "ego_fk_c2w": "ego_fk_c2w.npy (T,4,4) float64 camera-to-base pose of the ego left lens from the head FK (meta per_frame_cam_frame_left_wrt_base; R rows r0, r1, r0 x r1), row t = FK frame t. Nominal, NOT used for projection (cameras.npz uses the static T_ego_base fit).", |
| "time_base": "all labels are in the robot-log (FK) time base: video frame t shows log time t - 2.3 frames (tracks.time_offset_frames)." |
| }, |
| "stats": { |
| "visible_fraction": { |
| "view0": 0.4893, |
| "view1": 0.393 |
| }, |
| "in_frustum_fraction": { |
| "view0": 0.6749, |
| "view1": 0.5233 |
| }, |
| "never_visible_points": 282, |
| "visible_any_view_at_t16": 432, |
| "side_sign_ok": true, |
| "dup_max_um": 2.369, |
| "arm_block_rigidity_max_um": 0.21 |
| }, |
| "flag_rules": { |
| "calibration_step_exo": "episodes 000-022: the exo camera's frame 0 sits ~1.8 px lower than in episodes 023-099 with identical intrinsics / extrinsics (~2 px extrinsic error)", |
| "ego_extrinsic_manual_fit": "every episode: the ego (head) camera extrinsic is the source dataset's manual RGB fit (silhouette IoU ~0.8)", |
| "time_offset_residual_gt_0.5": "time_offset_residual_gt_0.5:<view>: the measured video lag of that view (attributes.measured_lag_frames) differs from the global -2.3 frames by more than 0.5 frame, with a lag-estimate peak correlation >= 0.8", |
| "few_default_queries_at_ref_frame": "few_default_queries_at_ref_frame:<n>: fewer than 20 points with default_gt visible in >= 1 view at frame 16 (n given)" |
| } |
| }, |
| "flags": [ |
| "ego_extrinsic_manual_fit" |
| ], |
| "files_sha256": { |
| "cameras.npz": "cbc96cca445466802918f16ad72128088df5164e612cb5e404594211a02b857c", |
| "captions.json": "ec34237375ff94337bc14b01f31a5f247629a7e84cae5fbc185dccf554b6f0b1", |
| "labels/action.parquet": "61c3bd425b5911528ae3baed78a17d729184ae60f46720f68ca607ca961e146f", |
| "labels/ego_fk_c2w.npy": "08518cc572702b56087f66aae02b0529ab4be74ab468b4847cf840bcad31622f", |
| "labels/proprio_30hz.parquet": "72a6dd51abe9bb7190eaf68c6389f082a5610ca30283e6df3d9f45a761b1bfcc", |
| "source/hand.npz": "4b1b7e517e86654288cdb243ebbbfdc9599fdd6aa7bec1fa6c6167806df1d386", |
| "source/meta.json": "c678c787825f1289813b1343c4988a1c35edad9baebf68013cafb636565c47d9", |
| "source/robot.npz": "e3feefb7ca8ce8f404bd64482998832f7da4f5891e5de9c5a2742e1706dc43ce", |
| "tracks3d.npz": "c49c7dc9cdb897da6934765cc659168971678accbc17abccb90ac111d40dd9a4", |
| "videos/view0.mp4": "3cf0a99a1fa0ebc090136e5d3ff9331a9bdee1a333d86489d4ce85ffdf22778c", |
| "videos/view1.mp4": "72c19d601a04461e077ba9ad8a59f249a17bad273a789f0e53edd512e17b5dd3" |
| } |
| } |
