fr3grasp pull-out-block demos (v1)
25 teleoperated demonstrations of a single task, "pull out the protruding block", recorded on a Franka FR3
with a Robotiq 2F-85 gripper. A GELLO leader arm drove a joint-impedance tracker (franky 2.0, libfranka 0.21)
that streams absolute joint targets. The data is in LeRobot v3.0 format. Camera streams are AV1 videos in videos/<camera>/chunk-000/file-000.mp4. States and actions are in data/chunk-000/file-000.parquet, and episode boundaries are in meta/episodes/. Load it with lerobot>=0.4.
- 25 episodes, all labelled successful by the operator; 8,567 frames at 15 fps
- 5 block placements (layouts) with 4–9 demos each; per-episode layout ids are in
meta/fr3grasp_episodes.jsonl - 2 cameras, both left eye at 224×224: exterior ZED 2i and wrist-mounted ZED Mini
- Language instruction:
pull out the protruding block
Columns
| column | shape / dtype | meaning |
|---|---|---|
image |
224×224×3 video | exterior camera, 1280×720 left eye letterboxed to 224×224 (scaled to 224×126, black bars above and below) |
extra_view_image |
224×224×3 video | wrist camera, same geometry |
state |
float32 [8] |
[gripper, q1..q7]: gripper = Robotiq position / 255 (0 = open, ≈0.65 when holding the block); q = measured joint angles (rad) |
actions |
float32 [8] |
7 joint velocities in normalized units, clipped to [-1, 1], followed by the gripper command in {0, 1} (1 = close) |
task_index |
int64 | always 0 |
episode_index, frame_index, index, timestamp |
standard LeRobot fields |
Action units
The joint-velocity actions integrate to the commanded joint target:
q_cmd[t+1] = q_cmd[t] + K * actions[t, :7]
K = [0.0404, 0.0283, 0.0408, 0.0291, 0.0379, 0.0447, 0.0415] # rad per step (1/15 s), per joint
q_cmd is the rate-limited joint target the controller received. When a step's velocity exceeds the [-1, 1]
range, it is clipped and the remainder is carried into the next steps, so the integral stays on the recorded
target. To deploy a policy trained on this data, integrate its actions the same way and send the result as an
absolute joint target. meta/fr3grasp_export.json holds K, the frame rate, the image geometry and the gripper
threshold.
Episode boundaries
Each episode starts once the leader arm is engaged and the robot has caught up to it; the initial alignment ramp is removed. Each episode ends when the operator presses Finish. The episodes contain no idle segments.
Extras
meta/fr3grasp_extra/episode_*.npz holds, for each episode, the commanded targets q_cmd, the measured joints
q and the sample times t, aligned frame by frame with the dataset frames.
The videos are lossy (AV1, about 42 dB PSNR against the source frames). A lossless copy with PNG frames in parquet is kept in LeRobot v2.1 format.
- Downloads last month
- 155