| "description": "Carnegie Mellon University and Bosch Center for Artificial Intelligence. Contact: yarun@andrew.cmu.edu. Data license: Creative Commons Attribution 4.0 International (CC BY 4.0); see LICENSE-DATA.txt and notes/license_note.json.\n\nDataset contents: 871 complete demonstration episodes across six tasks: 760 DFTP episodes (162 book, 208 insert_t, 163 scoop, 77 tea and 150 towel) and 111 valid-tactile F1 lab episodes. F1 recordings without informative tactile data are not part of this release. Original source split labels are preserved: 767 training, 57 validation and 47 unspecified episodes. Unspecified episodes have not been assigned to a training split. The lab instruction is Clean the lab. The 998,805 original rows yield 994,778 retained frames after removing 4,027 duplicate-timestamp rows; episode identities and row-selection provenance accompany this package.\n\nNative camera quality is preserved. The head stream selects the saved display-left stereo eye, assigned to the physical left eye from documented stereo evidence, and losslessly encodes the decoded RGB pixels with JPEG XL. Wrist streams retain the original JPEG payloads. Tea has only a head stream; other tasks also have both wrist streams. The training adapter resizes to 224x224 with padding; the stored images are not resized.\n\nInspireHandDFTP surface tactile is piezoresistive (owner-provided technology identification), with 1062 measured taxels per hand in 17 native patches. The saved dimensionless counts/4095 representation is preserved. Type=matrix groups pack same-area patches into 11 unique MTTS sites: six 15x8 maps in areas [4,3,2,1,0,6], four 10x8 maps in areas [14,13,12,11] and one 14x8 palm map in area 5. Each 3x3 patch is padded to 3x8 then stacked with its same-area 12x8 patch. Ninety synthetic zero cells per hand are described by masks and patch-placement metadata; the upstream tactile encoder does not automatically apply those masks.\n\nInspireHandF1 surface tactile is capacitive. Type=state preserves 37 source features per hand: five fingertips, each [normal_force_N/20, tangential_force_N/20, direction_sin, direction_cos, proximity], mapped little-to-thumb to MTTS [4,3,2,1,0]. Three palm regions, each with the first four features and no proximity, are concatenated into one 12-feature area 5 site. Proximity is a normalized clipped raw count, not metric distance. The direction pair (0,0) preserves the SDK no-contact flag encoded from direction=-1; nonzero force may coexist, so the flag is not proof of physical no contact. Exact direction axes and individual palm-region positions are not calibrated. Each feature is documented in notes/sensor_descriptions.json.\n\nBoth models export six additional motor-force feedback channels per hand, separate from surface tactile. Type=state groups map little/ring/middle/index to MTTS areas 20/19/18/17, and combine thumb flexion/rotation as two features in area 16. Values remain the saved force_act/1000 vendor scale; they are not calibrated joint torques in N m. There is no recorded wrist force/torque stream for area 15.\n\nRobot state includes seven arm joints per side, six hand joints per side in FAAS slots [1,2,7,12,17,22], and fifteen measured leg/waist supplementary joints, in recorded radians. supplementary_joints_note.json specifies all fifteen dimensions. FTP derives action targets from future measured states. Original commands, including forward/lateral/yaw velocity commands, remain unchanged in provenance/source_streams/actions; current FTP inputs/targets do not consume those archived commands.\n\nPose convention: XYZ metres plus rotation-vector radians in float64[T,6]. Recorded eef_to_body observations supply the wrists; fixed transforms return to the wrist origin and align the OXT canonical wrist axes. The standard left_wrist_pose/right_wrist_pose and camera_ego_pose share the moving torso_link reference. Explicit exact-value aliases are left_wrist_to_torso_pose, right_wrist_to_torso_pose and camera_ego_to_torso_pose_estimated. The auxiliary torso_pose_to_robot_ground_frame comes from recorded body_proprio in an instantaneous feet-defined kinematic frame. Neither reference is tracked room/world odometry. The constant camera mount causes FTP to mask head-motion targets. This torso-relative convention supports body-relative manipulation; it does not fulfill the tutorial wording of world-frame wrist trajectories.\n\nThe physical-left camera mount is estimated from supplied CAD, video and mounting photo: torso position [0.09081122829004291,0.030000000000002264,0.42294741448019024] m and 35.74886115349685 degrees downward. Camera extrinsics and intrinsics are estimates, not calibration. See notes/pose_and_camera.json.\n\nTimestamps use the recorded host camera-read times. Duplicate timestamps keep the last complete row consistently across exported modalities; all original numeric/Boolean rows, including discarded rows, remain in provenance. No files in the original or repaired source trees are modified. Per-episode source and output hashes are included. The earlier 20-episode official structural review retained a tea episode with a 0.101342-second timestamp interval, reported as a warning without interpolation. That historical result is scoped to the reviewed sample; the new full-release structural review is documented separately in quality_check_scope.json.\n\nQuality evidence has separate full-release and sampled-review scopes. All 871 episodes completed the converter's built-in source, shape, image-pixel and integrity checks, and each archive received member SHA-256, size and CRC verification during publication to the local export. A fresh audit also verified the SHA-256 hashes of all 871 archives (387,894,424,213 bytes), checked 1,685,079,992 numeric scalar values across all episodes, and confirmed informative force/proximity signals in both hands of all 111 included F1 episodes. The unchanged official OXT metadata/structure checker was run on all 871 episodes through a read-only ZIP discovery/storage adapter: 769 episodes have no warnings, 102 have timestamp-gap warnings, and none has a structural failure. All 165 flagged gaps are preserved from archived source timestamps, including eight intervals of approximately two seconds (maximum 2.070998 seconds); none was introduced or interpolated by conversion. Full outputs and the manual warning review are documented in quality_check_scope.json and notes/timing_warnings.json. The previously reviewed 20-episode example supplies independent source/RGB/tactile/pose validation, official sampled structural checks, actual-export 4D visualization, numerical signal diagnostics and assistant multimodal review of official OXT-generated RGB/tactile montages. Those 20 exports passed independent preservation validation, including 104,746 exact image comparisons. Actual FTP preparation and batching passed for those 20 episodes; normalization used their 26,570 training anchors and no validation data. All 180 bilateral surface/motor groups in that sample received visual review. These figures do not describe additional exhaustive visual review or model training on all 871 episodes. Contact and timing uncertainty remains documented, including the effect of negative plot-normalization minima that can render zero raw state features as long bars. The semantic review method and sampled-row coverage remain explicit; no external VLM API execution or model-training result is claimed. The full release is prepared locally for https://huggingface.co/datasets/chrisyrniu/HTD. Upload completion has not yet been verified.", |
