SleepMastger/dish-utensil-manipulation

Dish-utensil manipulation — real Franka teleoperation

100 human teleoperation demonstrations on a Franka Emika Panda: pick a utensil out of a drying rack and place it on a plate. Two utensils, 50 episodes each.

folder	utensil	episodes	sessions	collected
fork/	fork	50	6	2026-08-11
spoon/	spoon	50	4	2026-08-12

Layout is <utensil>/<session>/episode_<n>.hdf5, where the session is the collection timestamp (0811_1343 = Aug 11, 13:43). Sessions sort chronologically, so fork/0811_1146/episode_0.hdf5 is the first demonstration recorded and spoon/0812_1511/… the last.

Format

One HDF5 per episode, recorded at 10 Hz on a Franka under polymetis Cartesian impedance control with two RealSense cameras (scene + wrist).

obs/agentview_image    (T, 256, 256, 3) uint8   scene camera, RGB
obs/wrist_image        (T, 256, 256, 3) uint8   wrist camera, RGB
obs/gripper_width      (T, 1)  float32          metres, ~0.0796 open
obs/robot0_eef_pos     (T, 3)  float32          metres, robot base frame
obs/robot0_eef_quat    (T, 4)  float32          xyzw — SCALAR LAST
obs/robot0_joint_pos   (T, 7)  float32
actions                (T, 7)  float32

T varies per episode (roughly 300–550 steps, i.e. 30–55 s).

Actions

[dx, dy, dz, drx, dry, drz, gripper]

• dx, dy, dz — end-effector translation delta in metres per 100 ms step.
• drx, dry, drz — rotation deltas. The teleop rig commanded translation only, so these are identically zero in every frame.
• gripper — {-1, +1}, the raw value the collection tool logged.

⚠️ Gripper polarity. These are raw collector values, not remapped to {0, 1}. On this rig +1 commands a close and -1 an open — you can confirm it directly from the data, since obs/gripper_width is included: the width collapses on the frames where actions[:,6] becomes +1. If you convert to a {0=close, 1=open} convention, the usual mapping is g' = (1 - g) / 2.

Doing this check yourself is worth the two minutes. A sibling dataset that dropped gripper_width during a schema rewrite left its polarity unverifiable, which propagated into several trained checkpoints.

Observations

Both cameras are 256×256 RGB, stored exactly as the collector wrote them — no rotation, flip, or crop applied afterwards. obs[t] is the observation the operator saw when choosing actions[t].

What was removed

The raw recordings also carried 480×640 RGB, 480×640 and 256×256 depth, and a wrist IR stereo pair. Those are not included here: they were ~91% of the bytes and nothing in our training pipelines reads them. The channels above are the complete set consumed by the policies trained on this data.

Loading

import h5py, glob

for path in sorted(glob.glob("fork/*/episode_*.hdf5")):
    with h5py.File(path, "r") as f:
        rgb     = f["obs/agentview_image"][:]   # (T, 256, 256, 3) uint8
        wrist   = f["obs/wrist_image"][:]
        state   = f["obs/robot0_eef_pos"][:]    # (T, 3)
        actions = f["actions"][:]               # (T, 7)

Collection

Recorded with the tooling at github.com/sleepmastergx/franka-robot-tools (recorder/collect.py), teleoperated with a 3Dconnexion SpaceMouse. Every episode starts from the same saved home pose, so the initial frames are consistent within a session.