SleepMastger/pusht-manipulation

Push-T — real Franka teleoperation

100 human teleoperation demonstrations on a Franka Emika Panda: push an orange T-shaped block across a table until it aligns with a pink T outline marked on the surface. The gripper is clamped shut on a marker pen, which acts as a single-point pusher — this is the classic Push-T task run on real hardware rather than in simulation.

folder	episodes	frames	collected
0826_1637/	50	19,392	2026-08-26
0827_1114/	2	585	2026-08-27
0827_1118/	33	8,837	2026-08-27
0827_1144/	15	3,317	2026-08-27
total	100	32,131

Layout is <session>/episode_<n>.hdf5, where the session is the collection timestamp (0826_1637 = Aug 26, 16:37). Sessions sort chronologically, so 0826_1637/episode_0.hdf5 is from the first batch recorded. Episode numbers restart within each session and are not globally unique.

At 10 Hz the set is ~54 minutes of demonstration; episodes run 100–767 steps (10–77 s), median 292 (29 s).

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
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

Actions

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

• dx, dy, dz — end-effector translation delta in metres per 100 ms step, within ±0.018 m.
• drx, dry, drz — rotation deltas.
• gripper — open/close command.

⚠️ Four of the seven action channels are constant. The teleop rig commanded translation only, so drx, dry, drz are identically zero in every frame of every episode. And because the pen stays clamped for the whole task, gripper is identically +1 (closed) in all 32,131 frames — obs/gripper_width is likewise pinned at 0.0076 m throughout.

Only dx, dy, dz carry information. A min/max or standard-score normalizer fitted per channel will divide by zero on channels 3–6, which typically surfaces as NaNs partway into training rather than as an error at load time. Either restrict normalization to the first three channels, or train on a 3-dimensional action space and hold the rest fixed at their constant values.

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].

The end-effector stays within x ∈ [0.395, 0.606], y ∈ [−0.130, 0.312], z ∈ [0.221, 0.389] m in the robot base frame. It is not a strictly planar task: the 168 mm of z range is the operator lifting the pen between pushes to reposition without dragging the block.

Starting poses

99 of the 100 episodes begin from essentially the same pose — the median episode starts 4.8 mm from the median start, p90 6.3 mm, worst case 7.1 mm.

0827_1144/episode_13.hdf5 is the exception. Recording began with the arm parked elsewhere, and the operator spent the first three seconds driving it to the usual starting area before beginning the push. Those 30 lead-in frames have been trimmed from the episode published here, which is why it is 263 frames where the raw recording is 293; the file carries attrs["trimmed_lead_in_frames"] = 30, and it is the only episode in the set that has been edited.

Trimming reduces its offset from 292 mm to 26 mm, almost entirely a +287 mm y-axis discrepancy that is now gone. Note that 26 mm is still ~3.7x the worst of the other 99, so the episode is closer to the rest but not interchangeable with them — drop it if you need a strictly homogeneous initial-state distribution.

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 ~88% 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("*/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"][:, :3]           # only xyz carries signal

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.