cawm

Tactile manipulation data for world-model training: real-robot teleoperation on a Franka Panda arm, plus matched simulation rollouts. All streams are 10 Hz.

Preview

Each clip is wrist camera · agentview camera (where one was recorded) · FlexiTac contact map, played at 9.6× real time — the data is 10 Hz, so these are every 8th frame at 12 fps.

These are excerpts, not whole trajectories: a contiguous window of a single episode, chosen automatically as the stretch with the most tactile contact. Contact is sparse (under 10% of pad cells, and many frames have none at all), so a uniformly sampled window would mostly show a black tactile panel. The tactile panel is the merged (24, 32) grid used for training — on the single-pad rigs the upper and lower halves are duplicates of the one physical pad.

Real robot

Franka Panda, teleoperated at 10 Hz.

real/rubik

Rubik's cube manipulation. Wrist + agentview + one FlexiTac pad.
50 episodes · 100,000 steps · 166.7 min · T = 2000

real/allen_insert

Allen-key insertion, single recording session.
80 episodes · 112,792 steps · 188 min · T = 587–2797

real/erase_whiteboard

Whiteboard erasing — sustained sliding contact.
57 episodes · 90,708 steps · 151.2 min · T = 954–2332

real/usb_insertion

USB-connector insertion. See the caveats section below before training.
23 episodes · 62,265 steps · 103.8 min · T = 578–5000

Simulation

sim/rubik/hdf5/easy

Cube task, easy randomisation.
500 episodes · 199,500 steps · 332.5 min · T = 399

sim/rubik/hdf5/medium

Cube task, medium randomisation.
500 episodes · 199,500 steps · 332.5 min · T = 399

sim/rubik/hdf5/hard

Cube task, hard randomisation.
900 episodes · 359,100 steps · 598.5 min · T = 399

sim/rubik/hdf5/easy + sim/rubik/lerobot/easy

Two-camera build of the easy split: the agentview stream is recovered from the LeRobot copy and merged onto the HDF5 episodes.
500 episodes · 199,500 steps · 332.5 min · T = 399

sim/pen_in_hole/hdf5/all

Peg-in-hole insertion; variable episode length.
1000 episodes · 241,385 steps · 402.3 min · T = 125–299

---

real/usb_insertion

23 episodes of a USB-connector insertion task, teleoperated with a 3Dconnexion SpaceMouse and recorded at 10 Hz.

episodes	23
steps	62,265
duration	103.8 min
rate	10 Hz
format	HDF5, one file per episode (episode_0.hdf5 … episode_22.hdf5)

Contents of each file

obs/robot0_eef_pos                      (N, 3)          float32   measured EE position (m)
obs/robot0_eef_quat                     (N, 4)          float32   measured EE orientation, xyzw
obs/robot0_joint_pos                    (N, 7)          float32   joint angles (rad)
obs/gripper_width                       (N, 1)          float32   Franka Hand finger separation (m)
obs/images/cameras_agentview_rgb        (N, 224, 224, 3) uint8    fixed scene camera (RealSense D435i)
obs/images/cameras_wrist_rgb            (N, 224, 224, 3) uint8    wrist camera (RealSense D405)
obs/tactile/tactile_force_grid_left     (N, 16, 32)     float32   FlexiTac contact map, normalised
obs/tactile/tactile_force_grid_right    (N, 16, 32)     float32   see caveat below
actions                                 (N, 7)          float32   see below

File attributes: num_samples, action_dim, init_eef_pos (3,), init_eef_quat (4,).

Action space

7-dimensional, commanded deltas rather than achieved motion:

index	meaning
[0:3]	delta EE position, metres, base frame
[3:6]	delta EE rotation, axis-angle rotvec, radians, base frame (extrinsic)
[6]	gripper: +1 = closed, -1 = open

Observations are captured before the action is applied, so each (obs, action) pair is the state the operator saw and the command they gave in response. The commanded trajectory is init_eef_pos + cumsum(actions[:, :3]); the measured pose is obs/robot0_eef_pos, and the two differ where the impedance controller lagged the target (max 2.4–3.3 cm across these episodes).

Images are stored at the cropped 224×224 resolution used for training. The full uncropped sensor frames were never recorded.

Caveats — please read before training

Several dimensions are constant by construction and carry no signal:

• actions[3:6] is exactly zero in every step of every episode. Orientation was locked during collection, so the arm holds the orientation it homed at for the whole episode. This is a translation-only dataset.
• actions[6] is +1.0 everywhere. The gripper closes before recording starts and never reopens, so no grasp or release event appears in the data.
• gripper_width is ~0 throughout (−0.007 … 0.009 mm). The connector is held by a printed fixture rather than between the fingertips, so the fingers meet.
• tactile_force_grid_right is a byte-exact duplicate of ..._left. The rig has one FlexiTac pad; the second key was added so the schema matches a two-finger setup. It contains no independent measurement.

|dpos| reaches at most 6.00 mm per step (0.06 m/s at 10 Hz).

Provenance and validation

Merged from three collection sessions (2026-08-13 and 2026-08-19); real/usb_insertion/merge_manifest.txt maps each episode back to its source file.

Every episode was checked frame by frame before upload: no duplicate/stalled camera frames, no NaN/inf, consistent dataset lengths, and an identical schema across all 23 files.

License

Not specified by the authors. Please contact the repository owner before use.