tracyhan816/mug-video-touch-lora-hf25

Prepared training set · video + glove-touch LoRAs on the video model W (HF25)

The 25 training pairs of the public release tracyhan816/mug-sim2video-pose-tactile (commit 6cf129c), packaged exactly as the training scripts consume them, plus the held-out test cases. It is the data for the experiment experiments/finetune/2026-10-24 of github.com/tracyhann/tac-sim2real (branch for-amir): three LoRAs on the video model W (the backbone is anonymised for an internal test) that generate a 5 s video of a hand carrying a mug together with the TouchScale glove's 880-channel touch, differing only in the control video (FHN: scene depth + simulated hand / mug; SVO: simulation render only; CDS: cube-scene depth + simulation render). setup.sh in that repository downloads this dataset; nothing has to be built.

Contents (70 MB)

metadata_FHN.json, metadata_SVO.json, metadata_CDS.json   25 rows each, identical except control_file_path (one file per arm)
rows.json                                                 the same rows with all three control paths
video/<id>.mp4           832 x 480, 81 frames, 16 fps - the released video, unchanged (a zero-shot generation by the video
                         model W from a counterfactual physics replay of one real clip)
start/<id>.png           frame 0 of the video (the known first frame at training and inference)
tactile/<id>.npz         touch   (81, 880) float32  simulated normal force per glove sensor, newtons, at the video frames
                         S       scalar             the row's touch scale: median total force over frames with total > 1 N
                         touch_rel = touch / S      relative touch (what the model learns; absolute force is not identifiable from RGB)
                         u = min(touch_rel, 0.11) / 0.11   the touch head's target in [0, 1]
                         t       (81,)              frame times (s); sim_covered (880,) bool: 17 glove channels the hand model has no sensor for
handpose/<id>.npz        the simulated hand pose at the video frames (same file as hand_pose.npz of the release): kp_world / kp_cam /
                         kp_px (81, 21, 3|2) the 21 hand keypoints (kp_names, edges) in world, camera and pixel coordinates of the
                         832 x 480 video, in_image, wrist_T_world / wrist_T_cam (81, 4, 4), finger_q (81, 22) joint angles (finger_q_names),
                         base_q (81, 6), mug_T_world / mug_T_cam (81, 4, 4), T_wc (81, 4, 4) camera pose per frame, K, image_size, t_video, t_sim
action/<id>.npz          action-model view of the row (2026-10-25): state (81, 31) = wrist position xyz in the desk frame + wrist rotation
                         as rot6d + 22 finger joint angles (state_names); tactile (81, 880) N above baseline, tactile_rel = / S, tactile_u;
                         t, fps, S, schema. A policy predicts a chunk of future states relative to the chunk's first frame; the tactile
                         change over the chunk is the auxiliary target (see metadata_vtla.json "schema")
metadata_vtla.json       one manifest with every modality per row (video, start, handpose, action, tactile, controls, text, params)
PROMPTS.md               the table of all 25 prompts and parameters
ctrl/<clip>_<views>_<view>_{fun,sim}.mp4   control videos, 832 x 480 x 81: *_fun = inverse-depth grey of the static scene + the
                         simulated hand / mug / coaster rendered on top (FHN / CDS); *_sim = the simulated hand / mug / coaster alone
                         on black (SVO)
test_cases.json          7 evaluation cases: 4 held-out (two other real clips, never in the training rows, each from its real frame 0 +
                         real glove touch and from a zero-shot generated frame 0 + simulated touch) and 3 in-sample training rows;
                         each with the start image, the touch file and a control per arm
test/start/*.png, test/tactile/*.npz    the held-out starts and touch (same npz keys; the real rows' S is the real episode's scale)
test/simtouch/tsmug-*.npz               the held-out clips' own simulated touch (reference row "twin sim touch")
check.json               the build's per-row checks (frames, timing, touch range, control coverage)

Row fields: id, file_path, control_file_path, text (prompt v2: the shared task instruction "Pick up the ceramic mug from its coaster, lift it, carry it a short way, set it down and let go." + ONE quantitative clause for the row's perturbation or grasp, e.g. "While carrying it, shake the mug side to side, about 6 cm at 4 Hz."), params (the simulation's settings behind that clause: grip_scale, friction_scale, water_g, speed_scale, shake_amp_mm, shake_hz, posture, variant), type, width, height, group (r6 / r8b / r10), variant (recorded, grip030, fric05, water300, t10, shake60, claw, rimpinch), perturbation, camera, S, tactile, hand_pose, action, start, ctrl_own, ctrl_sim, ctrl_cube. Prompts v1 (until 2026-10-24) were a qualitative feature sentence plus a touch-panel sentence; the FHN LoRA runs of 10-24 used v1.

Groups: r6 = head camera, real first frame, scene depth from the fused real scan; r8b = head camera and a top view, cube-scene depth; r10 = user-supplied first frames (top views, a low side view, a front view, a cafe close-up), cube scene from the frame. Note: the 19 r8b / r10 videos were generated from cube controls, so for those rows the FHN and CDS controls are the same file.

Provenance

Videos and touch: the public release above (CC BY-NC 4.0). Controls: MuJoCo renders of the physics replays of the real clips (TouchScale glove recordings of a person carrying a mug), rendered without shadows into the camera of each video; the depth is either the fused real scan of the desk or a cube approximation of it. The held-out real touch is the glove's measured normal force above each sensor's resting baseline.

License

CC BY-NC 4.0, as the release it is built from.